Chap 03

3.
Lattice-gas cellular automata

3.1 The HPP lattice-gas cellular automata
The rst lattice-gas cellular automata (LGCA) was proposed in 1973 by
Hardy, de Pazzis and Pomeau. It is named HPP after the initials of the three
authors. The HPP model is the simplest
1
LGCA which will be discussed in
some length. Today HPP is of interest mainly for historical reasons
2
because
it does not lead to the Navier-Stokes equation in the macroscopic limit. In the
current chapter specic coding techniques for lattice-gas cellular automata
3
like multi-spin coding will be introduced.
3.1.1 Model description
HPP is a two-dimensional lattice-gas cellular automata model over a square
lattice. The vectors c
i
(i = 1, 2, 3, 4) connecting nearest neighbors (compare
Fig. 3.1.1) are called lattice vectors or lattice velocities. More precisely, the
lattice velocities are given by the lattice vectors divided by the time step
t which is always set equal to 1. So lattice vectors and lattice velocities
have dierent dimensions but the same numerical values. The meaning of the
c
i
can be easily recognized from the respective context. At each site (node)
there are four cells (Fig. 3.1.2) each associated to a link with the nearest
neighbor. These cells may be empty or occupied by at most one particle.
This exclusion principle (Pauli principle) is characteristic for all lattice-gas
cellular automata. It will lead to equilibrium distributions of Fermi-Dirac
4
1
The model of Boghosian and Levermore (1987) for Burgers equation in one
spatial dimension will not be discussed here.
2
There are still some applications: Chopard and Droz (1991) use HPP as a random
generator.
3
There are only few papers which give hints to specic coding techniques for
lattice-gas cellular automata; see, for example, Kohring (1991), Wolf-Gladrow
and Vogeler (1992), and Slone and Rodrigue (1997).
4
Fermi-Dirac distributions are well known from quantum mechanics. Particles
with half-odd-integer spins like the electron, proton, or neutron are called
fermions; they obey Fermi-Dirac distributions.
D.A. Wolf-Gladrow: LNM 1725, pp. 39138, 2000.
c Springer-Verlag Berlin Heidelberg 2000
40 3. Lattice-gas cellular automata
type for the mean occupation of the cells. All particles have the same mass
m (which will be set to 1 for simplicity) and are indistinguishable.
The evolution in time is deterministic and proceeds as an alternation of local
collisions c (only particles at the same node are involved) and streaming o
(also called propagation) along the appropriate links to the nearest neigh-
bors. The evolution operator c is dened as the composition of collision and
streaming:
c = o c. (3.1.1)
To each particle a momentum of magnitude mc
i
is assigned. The collision
should conserve mass and momentum while changing the occupation of the
cells. For HPP there is only one collision conguration. When two particles
enter a node from opposite directions and the other two cells are empty a
head-on collision takes place which rotates both particles by 90
in the same
sense (compare Fig. 3.1.3). All other conguration stay unchanged during
the collision step. In passing we note that twofold application of the collision
operator leads back to the initial conguration:
c
2
= 1, (3.1.2)
where 1 is the identity operator.
The HPP model respects a particle-hole symmetry, i.e. the operator T - which
interchanges particles and holes - commutes with the evolution operator c.
As a consequence of this symmetry the model has similar properties at low
and corresponding high mass densities (Hardy et al., 1973).
At each time step particles are interchanged between the sub-lattice consist-
ing of points with even indices (the white sub-lattice) and the sub-lattice
consisting of points with odd indices (the black sub-lattice; imagine a chess-
board). Therefore there exist two decoupled particle populations on the lat-
tice. This decoupling is characteristic for the square lattice (compare the
decoupling of solutions at even and odd time steps - sometimes called the
chessboard instability - in nite dierences; see, for example, Orszag, 1971 or
Rood, 1987).
As already noted above the HPP model does not obey the desired hydrody-
namic equations (Navier-Stokes) in the macroscopic limit. We will prove later
on that this decit is due to the insucient degree of rotational symmetry
of the lattice. Certain tensors composed of products of the lattice velocities -
so-called lattice tensors - are not isotropic over the square lattice. See Section
3.3 for an extensive discussion of these tensors. This anisotropy would man-
ifest itself, for example, in the ow past a non-rotational symmetric obstacle
in that the drag depends on the relative orientation of the obstacle with re-
spect to the lattice.
In addition to mass and momentum there exist additional conserved quanti-
ties for the HPP model. For example, the dierence in the number of parti-
cles parallel and anti-parallel to a lattice axis does not change by collisions
3.1 The HPP lattice-gas cellular automata 41
Fig. 3.1.1. The square lattice of the HPP model. The four arrows labelled by a,
b, c, and d indicate the lattice velocities c
i
. Particles are interchanged between the
black (small points) and the white (small circles) sub-lattices (chess board).
, , , , , , , , ,
, , , , , , , , ,
, , , , , , , , ,
, , , , , , , , ,
, , , , , , , , ,
, , , , , , , ,
, , , , , , , ,
, , , , , , , ,
, , , , , , , ,
Z
Z
`
``
Z
Z
`
`
a b
c d
y = 0
y = 1
y = 5
x = 0 x = 1
E
x
y
T
or propagation. These spurious invariants are undesirable because they re-
strict to a certain degree the dynamics of the model and have no counterpart
in the real world.
Fig. 3.1.2. HPP: collision and propagation. Filled circles denote occupied cells and
open circles empty cells. a) Part of the lattice before collision (e.g. after propaga-
tion); there is only one collision conguration (two particles in opposite cells at the
same node; on the left). b) After collision (e.g. before propagation): the conguration
of the cells at node on the left side has changed. c) After propagation: all particles
have moved along the links to their nearest neighbors (the lattice outside the part
shown was assumed to be empty, e.g. no propagation of particles from outside).
a) before collision
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z `
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
b) after collision
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z `
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
c) after propagation
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z `
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`

Fig. 3.1.3. HPP: collisions. a) There is only one collision conguration (head-on
collision) for HPP: two cell on opposite links are occupied and the two other cells
are empty. After collision the formerly empty cells are occupied and vice versa. b)
Same as a) but showing the associated momentum vectors. Both momentum vectors
are rotated by 90
. Mass and momentum are conserved.

a)
d
d
d
d
d
f
f
f
f

d
d
d
d
d
f
f
f
f
g
g
g
g
_
_
b)
d
d
d
d
d
d
d
d
`
``
`
`
Z
Z
Z
Z
_
_
3.1.2 Implementation of the HPP model: How to code lattice-gas
cellular automata?
The coding techniques discussed here will be also applicable to other LGCA
like FHP or PI. The FCHC model requires a dierent approach (discussed in
Section 3.5).
FORTRAN or C? Which programming language is best suited for the
coding of LGCA? Of course there will be no unique answer to this question
and often discussions with various people resemble religious controversies
5
. I
propose the following criteria:
The computer codes should be portable, i.e. programming in machine spe-
cic language (Assembler) is excluded.
The computation demand of hydrodynamic problems is usually large.
Therefore the model should be written in a language for which optimizing
compilers (vectorization, parallelization) are available on super-computers.
Thus BASIC or PASCAL are excluded and one has to choose between C
or FORTRAN.
The specic coding techniques for LGCA can be applied in C as well as
in FORTRAN. Comparison of computation time shows only a very small
advantage of C (Wolf-Gladrow and Vogeler, 1992). The code of the collisions is
much easier to grasp in C than in FORTRAN. Nevertheless, the translation
from C to FORTRAN is straightforward (compare Table 3.1.1).
Multi-spin coding. The most important technique for LGCA is multi-spin
coding
6
. The exclusion principle makes it possible to describe the state of a
cell by one bit which is set to 0 if the cell is empty and to 1 if it is occupied.
There is no special data type for bits either in C nor in FORTRAN. However,
several bits (32 on Sun-Workstation; 64 on CRAY-J90) can be packed into
one unsigned (C) or integer (FORTRAN) variable. In standard C there exist
bit-operators which act bitwise
7
on whole unsigned variables. In FORTRAN
you may nd bit-functions with the same eects. These bit-functions do not
belong to the FORTRAN standard but are available on almost all machines.
The core of the HPP program, namely the coding of collision and streaming,
encompasses only a few lines. Here is the code in C:
5
I was taught that FORTRAN is a dead language already in the 70ies.
6
Although the term multi-spin coding which is related to the spins of Ising mod-
els has been coined by Creutz et al. in 1979, this technique has been described
already in the article of Hardy et al. (1976) and it was rst mentioned by Fried-
berg and Cameron (1970).
7
Example: decimal 65 | 39 reads in binary notation on an 8-bit machine
(01000001) | (00010111) = (01010111) and therefore 65 | 39 = 103.
Table 3.1.1. Some elements and constructions in C and FORTRAN.
C FORTRAN Remarks
----------------------------------------------------------------
a & b iand(a,b) bit operators are
a | b ior(a,b) standard in C;
a b ieor(a,b) bit functions are
a not(a) (almost always available)
extensions in FORTRAN
a<<3 ishift(a,3) left shift of bits by
3 positions
a>>4 ishift(a,-4) right shift
a = a | (1<<3) a = ibset(a,4) set 4th bit
#define A 5 parameter (A=5) in C: global define
unsigned a[3] integer a(3) 1D array with 3 elements
a[0],a[1],a[2] a(1),a(2),a(3) elements
for(i=0;i<3;i++){ do i=1,3 loops
} enddo
a % b mod(a,b) a modulo b
----------------------------------------------------------------
Table 3.1.2. Bit-operators in C: & and; | inclusive or; exclusive or; not.
a b a & b a | b a b a
0 0 0 0 0 1
1 0 0 1 1 0
0 1 0 1 1 1
1 1 1 1 0 0
/* --- grid ---
b a
\ /
\ /
\ /
+
/ \
/ \
/ \
c d */
last = LENGTH - 1; /* --- shift last bit --- */
XMAX1 = XMAX - 1;
YMAX1 = YMAX - 1;
/* ----- collision ----- */
for(x=0; x<XMAX; x++)
for(y=0; y<YMAX; y++) {
change = ( (a1[x][y] & c1[x][y] & (b1[x][y] | d1[x][y])) |
(b1[x][y] & d1[x][y] & (a1[x][y] | c1[x][y])) )
& nsb[x][y];
a2[x][y] = a1[x][y] change;
b2[x][y] = b1[x][y] change;
c2[x][y] = c1[x][y] change;
d2[x][y] = d1[x][y] change; }
/* ----- propagation in a-direction ------ */
for(x=1; x < XMAX1; x++) {
for(y=1; y < YMAX1; y += 2) {
/* note: brackets are necessary,
because in C the + has a higher priority than the >> */
/* black to white: */
a1[x][y] = (a2[x][y-1] >> 1) + (a2[x-1][y-1] << last);
a1[x][y+1] = a2[x][y]; }} /* white to black: */
A few comments on the code are now in order:
Each unsigned variable can store LENGTH (= 64 on CRAY computers)
bits. The number of grid points is XMAX*LENGTH in x-direction and
YMAX in y-direction (compare Fig. 3.1.1).
The states of all cells are stored in the two-dimensional (2D) arrays a1,
b1, c1, d1, where a, b, c, d assign the dierent lattice directions (lattice
velocities). a1[0][0] contains the a-bits of the nodes from 1 to LENGTH
of the rst line.
The location of obstacles is stored in the 2D array nbs (non-solid bit): bits
in nbs are set to 1 outside of the obstacles and 0 otherwise.
In the rst loop a variable change is calculated. It contains the information
whether or not a collision will happen: the bits in change are set to 1 if
the congurations (a, b, c, d) = (1010) or (0101) are present and the node
is located outside of obstacles.
Subsequently, the bit arrays a1, b1, c1, d1 are concatenated by exclusive
or with the variable change. This changes the state bits (interchange of 0
and 1; compare Table 3.1.2). The results of this operation are stored in the
auxiliary arrays a2, b2, c2, d2. The introduction of these auxiliary arrays
ensures a very fast updating on vector computers. These arrays are also
useful in the propagation step.
In Fig. 3.1.1 the propagation is shown only for the a-directions (upwards
and to the right). In the program listing the propagation in a-direction is
shown only for the inner nodes. The propagation for nodes on the bound-
aries has to be treated separately according to the appropriate boundary
conditions.
The propagation from the white to the black sub-lattice consists of a
storage in dierent arrays.
The propagation from the black to the white sub-lattice is more involved.
To simplify the following discussion let us assume that number of bits per
integer is only 4 (the parameter LENGTH in the code). The propagation
of a2[0][5] and a2[1][5] to a1[1][6] (compare Fig. 3.1.1) will be considered.
First all bits of a2[1][5] will be shifted to the right whereby the rightmost
bit drops out. The resulting void at the left boundary of a2[1][5] will be
lled up automatically by a 0. Yet, this position must be occupied by the
rightmost bit of the neighbor element a2[0][5]. To isolate this bit all bits
in a2[0][5] are shifted to the left by LENGTH-1 (= last in the code)
digits whereby zeros ll up from the right. Now a2[1][5] after a right
shift and a2[0][5] after LENGTH-1 left shifts could be concatenated by
exclusive or. But the addition of these two shifted integers yields the
same result and is faster on some computers due to chaining
8
(compare
Kohring, 1991 and Wolf-Gladrow and Vogeler, 1992).
Example: LENGTH = 4; last = 3;
8
Chaining is the process of passing the output of one vector operation directly
as input into another vector operation. As soon as the rst element of the rst
operations result is output, the second operation can begin. This allows partial
overlapping of vector instruction execution.
a2[0][5] = (1001); a2[1][5] = (1011);
a2[1][5] >> 1 = (0101); a2[0][5] << last = (1000)
a1[1][6] = (1101)
3.1.3 Initialization
Before the evolution of the LGCA can start the various arrays have to be
initialized. At time t = 0 the bits are set by random processes with prob-
abilities such that the mean values over a large number of nodes (typically
32 times 32 or 64 times 64) approximate the given initial values for mass
and momentum density. Thus the question arises, how to choose appropriate
probabilities for given mass and momentum density?
The state of the LGCA is fully described by the Boolean elds n
i
(t, r
j
) where
the index i which runs from 1 to 4 (or alternatively from a to d) indicates the
directions, n
i
is the occupation number which may be 0 or 1, t is the (discrete)
time and r
j
are the coordinates of the nodes. Mean occupation numbers N
i
are calculated by averaging over neighboring nodes
N
i
(t, x) = n
i
(t, r
j
) . (3.1.3)
The mean occupation numbers can take on values between 0 and 1. Mass
(t, x) and momentum density j(t, x) are dened by
(t, x) =
4
i=1
N
i
(t, x) (3.1.4)
and
j(t, x) = u =
4
i=1
c
i
N
i
(t, x) (3.1.5)
(u is the ow velocity) with the lattice velocities c
i
c
1
=
1
2
(1, 1) (3.1.6)
c
2
=
1
2
(1, 1)
c
3
=
1
2
(1, 1) (3.1.7)
c
4
=
1
2
(1, 1)
which obey
4
i=1
c
i
= 0 (lattice symmetry!) (3.1.8)
and
4
i=1
c
i
c
i
= 2
, (3.1.9)
where the Latin indices refer to the lattice vectors and run from 1 to 4
whereas the Greek indices assign the cartesian components of the vectors
and therefore run from 1 to 2. This convention will be used also in all other
chapters.
The theoretical background for the calculation of the equilibrium occupation
numbers will be developed not until the next chapter (the section on the FHP
model contains some results relevant for HPP). Instead a simple ansatz for
N
i
will be made which is linear in and j
N
i
= +c
i
j. (3.1.10)
The coecients and can be calculated from the constraints (3.1.4) and
(3.1.5):
=
4
i=1
N
i
=
i
+j
i
c
i
. .
= 0
= 4 (3.1.11)
which yields = 1/4;
j =
4
i=1
c
i
N
i
=
i
c
i
. .
= 0
+
i
c
i
(c
i
j)
= 2j (3.1.12)
thus = 1/2 and therefore
N
i
=

4
+
1
2
c
i
j. (3.1.13)
I.e., for j = 0 the occupation numbers are independent of direction (they can
still depend on location which corresponds to pure density perturbations)
whereas non-vanishing momenta imply occupation numbers which vary with
direction.
The Boolean arrays n
i
will be initialized with probabilities
9
p
i
such that the
N
i
give the desired distributions of and j when summed up according to
Eqs. (3.1.4) and (3.1.5). The relations between the Boolean arrays n
i
and the
mass and momentum density are illustrated in Fig. 3.1.4.
Fig. 3.1.4. Relations between microscopic (Boolean arrays n
i
) and macroscopic
(mass and momentum density) level.
_
_
n
i
. .
N
i
. .
j
..
averaging
throw dice denition
0, 1
Boolean
occupation number
IR
mean
occupation number
mass and
momentum
density
Exercise 3.1.1. (*)
Construct occupation numbers N
i
which vary with direction but yield j = 0.
Exercise 3.1.2. (**)
How long does it take for the distribution (3.1.13) to relax toward equilibrium
distribution? What does the relaxation time constant depend on?
3.1.4 Coarse graining
The calculation of mean values for mass and momentum density is called
coarse graining. Although it is possible to average over space, time or a com-
bination of space and time, spatial coarse graining is much faster than the
9
Unfortunately there is no standard for random generators. Portable random gen-
erators can be found for example in Numerical Recipes (Press et al., 1992a,b).
Dierent types of random generators (multiplicative congruential, shift-register
and lagged Fibonacci) are discussed by Slone and Rodrigue (1997).
other alternatives. For the purpose of coarse graining the domain is divided
into a number of subdomains which are large enough (usually 32 times 32 or
64 times 64 nodes) to obtain reliable (low noise) averages and small enough
as to allow a large physical domain under the constraint of a given limit of
core memory.
The following more technical notes can be skipped in a rst reading. The
main computational load is the counting of the 1-bits in the unsigned (in-
teger) arrays. On some computers a fast routine for counting the 1-bits in
an integer is available. On the CRAY, for example, this is the FORTRAN
function POPCNT (population count). Nothing similar is available in C. But
one can use the FORTRAN function in C. Include the following lines into
the code
#include <fortran.h>
fortran int POPCNT(); /* counts the number of 1-bits
in a 64 bit word */
and apply POPCNT to unsigned variables:
int n;
unsigned u;
u = 7;
n = POPCNT(u);
If no machine specic population count is available one can use the following
routine which applies a look-up table (see Kohring, 1991, for the FORTRAN
version):
/* ------- popcount for 32-bit unsigned ------ */
unsigned lu16[65536]; /* global */
void makelu16()
{
/* ----- make look-up table ----- */
int ilu,ibi;
for(ilu=0; ilu<65536; ilu++) {
lu16[ilu] = 0;
for(ibi=0; ibi<16; ibi++) {
if( (ilu & bits[ibi]) > 0) lu16[ilu] += 1;
}
}
} /* --- end of makelu16 --- */
int popcount(unsigned u)
{
/* --- count bits --- */
int pop;
pop = lu16[(u&65535)] + lu16[( (u>>16)&65535 )];
return(pop);
} /* --- end of popcount --- */
3.2 The FHP lattice-gas cellular automata 53
3.2 The FHP lattice-gas cellular automata
In 1986 Frisch, Hasslacher and Pomeau showed that a lattice-gas cellular
automata model over a lattice with a larger symmetry group than for the
square lattice yields the incompressible Navier-Stokes equation in the macro-
scopic limit. This model with hexagonal symmetry is named FHP according
to the initials of the three authors. The discovery of the symmetry constraint
was the start for a rapid development of lattice-gas methods. The theoretical
foundations where worked out by Wolfram (1986) and by Frisch et al. (1987).
Within the following years many extensions and generalizations (FCHC for
3D simulations, colored models for miscible and immiscible uids) were pro-
posed. These models allow a wide range of applications.
After concentrating more on coding techniques in the HPP chapter the focus
of the current section will be on the theory of lattice-gas cellular automata.
Especially the equilibrium distribution and the macroscopic equations will
be derived.
3.2.1 The lattice and the collision rules
The FHP lattice is composed of triangles (compare Fig. 3.2.1). It is invariant
under rotations by n 60
modulo 360
(hexagonal symmetry) about an axis

through a node and perpendicular to the lattice plane. At each node and each
link to the nearest neighbor there is a cell which may be empty or occupied by
at most one particle (exclusion principle). All particles have the same mass
m (set to 1 for simplicity) and are indistinguishable. The state of a node
can be described by six bits. The exclusion principle leads to an equilibrium
distribution of the mean occupation numbers of Fermi-Dirac type. Each cell
is associated with a lattice vector c
i
which connects a node with its nearest
neighbor in direction i. The lattice vectors c
i
are also called lattice velocities
because the time step t is always set to 1 in lattice-gas cellular automata
and therefore c
i
and c
i
/t have the same numerical values. Because all par-
ticles have the same mass m = 1, c
i
is also the particles momentum. The
respective meaning of c
i
can be recognized from the context.
As for HPP there are 2-particle head-on collisions (compare Fig. 3.2.2). The
initial state
10
(i, i + 3) can be transformed into one of two dierent nal
states (i +1, i +4) or (i 1, i +2) (rotation by 60
to the left or right) while

conserving mass and momentum density. If one chooses always one and the
10
The description of the state of a node will be given in terms of indices of the
occupied cells whereby the cell index i > 0 is understood as modulo 6. The index
i = 0 will be assigned to rest particles (see below).
same nal state the model becomes chiral
11
: it is not invariant with respect
to spatial reections (parity transformation). This is an undesired property
because the hydrodynamic equations do not break parity symmetry. To re-
store reection symmetry on the macroscopic level the choice between the
dierent nal states will be made by a random process with equal probabil-
ities for rotation to the left and to the right. Thus in contrast to HPP the
FHP model encompasses nondeterministic rules.
The generation of random numbers is a time consuming process. Therefore a
pseudo-random choice is used where the rotational sense changes by chance
for the whole domain from time step to time step (i.e. only one random num-
ber per time step has to be generated) or the sense of rotation changes from
node to node but is constant in time (i.e. random numbers have to be gener-
ated only in the initial step for all nodes).
The 2-particle collisions conserve not only mass and momentum but also the
dierence of the number of particles that stream in opposite directions (the
same invariant as for HPP). This additional invariant has no counterpart in
real world hydrodynamics and therefore is called a spurious invariant. It
further restricts the dynamic of the lattice-gas cellular automata and can lead
to deviations from hydrodynamic behavior on the macroscopic scale. The in-
variance of the particle dierences can be destroyed by symmetric 3-particle
collisions which conserve mass and momentum (compare Fig. 3.2.2). 2- and
3-particle collisions form a minimal set of collisions for FHP. This version
of FHP is called FHP-I. Introduction of additional collisions like 4-particle
collision, 2-particle collision with spectator and collisions including rest parti-
cles lead to various variants (FHP-II, FHP-III: see, for example, Frisch et al.,
1987 and Hayot and Lakshmi, 1989). The corresponding macroscopic equa-
tions all have the same form (universality theorem) and dier only in their
viscosity coecients. As a rule of thumb the viscosity coecient decreases
with increasing number of collisions.
The 3-particle collisions destroy a spurious invariant. Unfortunately no
method exists to detect all invariants of a given lattice-gas cellular automata
model. Of course this is an unsatisfactory situation in the light that the in-
variants play an essential role in the equilibrium distributions of the mean
occupation numbers. Indeed Zanetti (1989) found spurious invariants for all
variants of the FHP model. Fortunately these staggered invariants are not
set to values above a certain noise level by the usual initialization procedure
and obviously are not generated by interactions with obstacles. So they do
not inuence the macroscopic dynamic too much. A discussion of the Zanetti
invariants will be given in Section 3.8.
Frisch et al. (1987) give a discussion of a unied theory for the lattice-gas
cellular automata HPP, FHP and FCHC. This excellent paper may be heavy
fare for the beginner. The following discussion is restricted to the FHP model
11
The word chiral derives from the Greek word for hands.
Fig. 3.2.1. The triangular lattice of the FHP model shows hexagonal symmetry.
The lattice velocities c
i
are represented by arrows.
, , , , , , ,
, , , , , , ,
, , , , , , ,
, , , , , , , ,
, , , , , , , ,
E '

!
e
eu
e
e
E '
1
T
c
3
2
without rest particles (FHP-I) but from time to time more general results
from Frisch et al. (1987) will be quoted.
The essential properties of the FHP model read:
1. The underlying regular lattice shows hexagonal symmetry.
2. Nodes (also called sites) are linked to six nearest neighbors located all at
the same distance with respect to the central node.
3. The vectors c
i
linking nearest neighbor nodes are called lattice vectors or
lattice velocities
c
i
=
_
cos

3
i, sin

3
i
_
, i = 1, ..., 6. (3.2.1)
with [c
i
[ = 1 for all i.
4. A cell is associated with each link at all nodes.
5. Cells can be empty or occupied by at most one particle (exclusion prin-
ciple).
6. All particles have the same mass (set to 1 for simplicity) and are indis-
tinguishable.
7. The evolution in time proceeds by an alternation of collision c and
streaming o (also called propagation):
c = o c, (3.2.2)
where c is called the evolution operator.
8. The collisions are strictly local, i.e. only particles of a single node are
involved.
Fig. 3.2.2. All possible collisions of the FHP variants: occupied cells are represented
by arrows, empty cells by thin lines.
a) g _ _

e
e
e
e
p = 0.5
.
.
.
.
.
.,
-
-
-
-
-
-
2-particle head-on
collisions
g
/
/
/
/
e
e
e
e
g
`
`
`
`*
b) g _
/
/
`
`*
e
e
_
_
symmetric 3-particle
collisions
g _
/
/
`
`
e
e
c) g
/
/
/
/
`
`
`
`*
p = 0.5
.
.
.
.
.
.,
-
-
-
-
-
-
4-particle head-on
collisions
g _ _
`
`
`
`*
g _ _
/
/
/
/
e
e
e
e
d) g _ _
/
/
e
e
e
e
_
_
2-particle head-on
collisions with spectator
g
/
/
`
`
`
`*
e) _g _
e
e
e
e
_
_
rest particle (circle)
collisions
g
/
/
`
`*
e
e
r is the position vector of a node and r + c
i
are the position vectors of its
nearest neighbors. Each pair of lattice vectors (c
i
, c
j
) is associated with an
element of the isometric group
12
( which transforms c
i
into c
j
by a rotation
of n 60
. The rst three moments of the lattice velocities c

i
read:
i
c
i
= 0 (symmetry of the lattice!) (3.2.3)
i
c
i
c
i
= 3
(3.2.4)
i
c
i
c
i
c
i
= 0 (3.2.5)
where the sum over i always runs from 1 to 6. The Latin indices refer to
the lattice vectors and run from 1 to 6 whereas the Greek indices assign
the cartesian components of the vectors and therefore run from 1 to 2. To
each node a bit-state n(r) = n
i
(r), i = 1, ..., 6 will be assigned where the
n
i
0, 1 are Boolean variables. The streaming (propagation) is dened by
n
i
(r) = on
i
(r c
i
). (3.2.6)
Collisions take place synchronously at every node and transform the initial
state of a node s = s
i
, i = 1, ..., 6 into the nal state s
= s
i
, i = 1, ..., 6
according to the collision rules described above. A certain transition proba-
bility
A(s s
) 0 (3.2.7)
is assigned to each pair of initial and nal state. The transition probabilities
satisfy the normalization
s :
A(s s
) = 1 (3.2.8)
The only combinations of six real numbers a
i
which fulll the constraints
s, s
i
(s
i
s
i
) A(s s
) a
i
= 0 (3.2.9)
are linear combinations of 1 (for all i) and c
i
, i.e. mass and momentum
conservation. This is indeed the case for FHP. The Zanetti invariants are
non-local invariants.
The transition probabilities are invariant with respect to each element of the
isometric group (:
g (, s, s
: A(g(s) g(s
)) = A(s s
) . (3.2.10)
12
Isometries are mappings g which keep the distances d(, ) of arbitrary points
, invariant: d(g(), g()) = d(, ). Rotations and reections are isometries.
The FHP model fullls the detailed balance
A(s s
) = A(s
s), (3.2.11)
i.e. the probabilities for each collision and its inverse collision are equal. For
the derivation of the equilibrium distribution (compare Theorem 3.2.1) the
weaker condition of semi-detailed balance (or Stueckelberg condition
13
)
s
s
A(s s
) = 1 (3.2.12)
is sucient.
There are 2
6
= 64 dierent states for a node of the FHP-I model. The tran-
sition probabilities A(s s
) form a 64 64 transition matrix A

ss
. The two
constraints (3.2.8) and (3.2.12) are equivalent to the statement that the sums
of each line or each column of A
ss
are equal to 1.
Exercise 3.2.1. (*)
How does the transition matrix A
ss
look like? Write down the submatrix
which contains the 2- and 3-particle collisions.
Prove: the 2- and 3-particle collisions locally (at a single node) conserve only
mass and momentum.
Exercise 3.2.3. (*)
Consider a system with three possible states (a, b, c) and transition proba-
bilities
1.)
A(a b) = 0.5 A(b a) = 0.5
A(b c) = 0.5 A(c b) = 0.5
A(c a) = 0.5 A(a c) = 0.5
or 2.)
A(a b) = 0.8 A(b a) = 0.2
A(b c) = 0.8 A(c b) = 0.2
A(c a) = 0.8 A(a c) = 0.2
13
Stueckelberg (1952) showed that this condition is sucient to prove the H-
theorem.
or 3.)
A(a b) = 0.8 A(b a) = 0.3
A(b c) = 0.7 A(c b) = 0.4
A(c a) = 0.6 A(a c) = 0.2.
Calculate the distribution (a, b, c)
t
for large t given the initial value
(a, b, c)
t=0
= (1, 0, 0) for all three cases. In which cases is detailed or semi-
detailed balance fullled?
3.2.2 Microdynamics of the FHP model
For hydrodynamics certain averaged quantities like mass and momentum den-
sity are of interest. To understand the behavior of these averaged quantities a
description of lattice-gas cellular automata in terms of statistical mechanics
will be formulated. First the analogue to the Hamilton equations in classical
statistical mechanics will be considered.
Boolean description of the microdynamics. The state of the cells of the
lattice-gas cellular automata is described by the Boolean
14
arrays n
i
(t, r):
n
i
(t, r) =
_
1 if cell i is occupied
0 if cell i is empty
where r and t indicate the discrete points in space and time. The time evo-
lution in terms of Boolean arrays reads:
n
i
(t + 1, r +c
i
) = n
i
(t, r)
[(n
i
n
i+1
) &(n
i+1
n
i+2
) &(n
i+2
n
i+3
)
&(n
i+3
n
i+4
) &(n
i+4
n
i+5
)]
[ [n
i
&n
i+3
& (n
i+1
[ n
i+2
[ n
i+4
[ n
i+5
)] (3.2.13)
[ [&n
i+1
&n
i+4
& (n
i
[ n
i+2
[ n
i+3
[ n
i+5
)]
[ [ &n
i+2
&n
i+5
& (n
i
[ n
i+1
[ n
i+3
[ n
i+4
)]
where the occupation numbers n
i
on the right side refer to the input-states
at node r and time t. The symbols
14
A short introduction to the Boolean algebra can be found in Appendix 6.1.
& = AND
[ = OR (inclusive or) (3.2.14)
= XOR (exclusive or)
= NOT.
have been used. is a Boolean random variable which determines the sense of
rotation for the 2-particle collisions (denoted xi in the program code listed
below). Eq. (3.2.13) looks rather nasty to people who do not use Boolean
algebra every day. The interpretation is however simple. Eq. (3.2.13) states
that the state of cell i at node r + c
i
at time t + 1 is given by n
i
(r, t) when
no collision happens or (exclusive!) n
i
(r, t) is changed by collision before its
value is propagated to r +c
i
. The following collisions are taken into account:
1. A symmetric 3-particle collision happens at node r at time t when the
n
j
(j = 1, ..., 6) are alternating occupied and empty. When a 3-particle
collision happens and n
i
is occupied at input cell i will change to the
empty state, whereas in a 3-particle collision with n
i
empty initially cell
i will be occupied after collision.
2. If a 2-particle collision happens where on input cells i and i + 3 are
occupied (and all other cells empty) n
i
will change to 0.
3. If a 2-particle collision happens where on input cells i + 1 and i + 4 are
occupied (and all other cells empty) and the Boolean random variable
is 1 such that the rotation of the outgoing particles is in the clockwise
direction (otherwise n
i
would not change!).
4. If a 2-particle collision happens where on input cells i + 2 and i + 5 are
occupied (and all other cells empty) and the Boolean random variable
is 0 such that the rotation of the outgoing particles is in the counter-
clockwise direction.
The four dierent cases correspond to the last ve lines on the right hand side
of Eq. (3.2.13). n
i
is changed if any one of the collisions happens. Therefore
the various collision terms are connected by inclusive or. The exclusive or
between n
i
(r, t) and the collision terms enforces a change when any one of
the collisions happens.
The coding of the FHP model is based on the Boolean formulation of the
microdynamics. The code in C reads as follows (nsbit is the non-solid bit
which is set to 1 in the uid and to 0 inside obstacles; it is used here to
suppress collisions inside obstacles):
/* loop over all sites */
for(x=0; x<XMAX; x++) {
for(y=0; y<YMAX; y++) {
a = i1[x][y];
b = i2[x][y];
c = i3[x][y];
d = i4[x][y];
e = i5[x][y];
f = i6[x][y];
/* two-body collision
<-> particles in cells a (b,c) and d (e,f)
no particles in other cells
<-> db1 (db2,db3) = 1 */
db1 = (a&d&(b|c|e|f));
db2 = (b&e&(a|c|d|f));
db3 = (c&f&(a|b|d|e));
/* three-body collision <-> 0,1 (bits) alternating
<-> triple = 1 */
triple = (ab)&(bc)&(cd)&(de)&(ef);
/* change a and d
<-> three-body collision triple=1
or two-body collision db1=1
or two-body collision db2=1 and xi=1 (- rotation)
or two-body collision db3=1 and noxi=1 (+ rotation)
<-> chad=1 */
xi = irn[x][y]; /* random bits */
noxi = xi;
nsbit = nsb[x][y]; /* non solid bit */
cha = ((triple|db1|(xi&db2)|(noxi&db3))&nsbit);
chd = ((triple|db1|(xi&db2)|(noxi&db3))&nsbit);
chb = ((triple|db2|(xi&db3)|(noxi&db1))&nsbit);
che = ((triple|db2|(xi&db3)|(noxi&db1))&nsbit);
chc = ((triple|db3|(xi&db1)|(noxi&db2))&nsbit);
chf = ((triple|db3|(xi&db1)|(noxi&db2))&nsbit);
/* change: a = a chad */
k1[x][y] = i1[x][y]cha;
k2[x][y] = i2[x][y]chb;
k3[x][y] = i3[x][y]chc;
k4[x][y] = i4[x][y]chd;
k5[x][y] = i5[x][y]che;
k6[x][y] = i6[x][y]chf;
/* collision finished */
}}
Write the analogue of Eq. (3.2.13) for the case where rest particle collisions
(compare Fig. 3.2.2 e) are included.
Arithmetic description of the microdynamics. The transition from the
Boolean to the arithmetic formulation of the microdynamics can be achieved
by formal substitutions (compare Table 3.2.1). However, this procedure is too
laborious. Instead the collision function
i
is introduced by
n
i
(t + 1, r +c
i
) = n
i
(t, r) +
i
. (3.2.15)
The collision function can be constructed according to the following recipe:
For each collisional conguration, i.e. one that will lead to a change of the
occupation numbers, a product of the n
i
(if n
i
= 1) respectively (1 n
i
)
(if n
i
= 0) for i = 1, ..., b (b total number of lattice velocities) is written.
This product yields 1 if the specic conguration is given and 0 otherwise.
For nondeterministic collision rules (FHP) the products dened above are
multiplied by a random variable , (1 ) or 1, respectively.
The products receive positive or negative sign according to whether n
i
increases or decreases when changing from input (before collision) to output
(after collision) state.
All products are added up.
To give an example let us consider the HPP model. There are only two col-
lisional congurations, namely (1, 0, 1, 0) and (0, 1, 0, 1). The corresponding
products read
n
i+1
n
i+3
(1 n
i
)(1 n
i+2
) and n
i
n
i+2
(1 n
i+1
)(1 n
i+3
)
where n
i
increase in the former case and decreases in the latter. Thus the
collision function is given by
i
= n
i+1
n
i+3
(1 n
i
)(1 n
i+2
) n
i
n
i+2
(1 n
i+1
)(1 n
i+3
).
Table 3.2.1. Translation of Boolean expressions into arithmetic formulas.
a b
AND OR XOR NOT
a&b a | b a b a
a b a + b a b a + b 2 a b 1 a
0 0 0 0 0 1
1 0 0 1 1 0
0 1 0 1 1 1
1 1 1 1 0 0
The analogous procedure yields for the FHP-I model
i
(n) = n
i+1
n
i+3
n
i+5
(1 n
i
)(1 n
i+2
)(1 n
i+4
)
n
i
n
i+2
n
i+4
(1 n
i+1
)(1 n
i+3
)(1 n
i+5
)
+n
i+1
n
i+4
(1 n
i
)(1 n
i+2
)(1 n
i+3
)(1 n
i+5
) (3.2.16)
+(1 )n
i+2
n
i+5
(1 n
i
)(1 n
i+1
)(1 n
i+3
)(1 n
i+4
)
n
i
n
i+3
(1 n
i+1
)(1 n
i+2
)(1 n
i+4
)(1 n
i+5
)
where n = n
i
, i = 1, ..., 6. The conservation of mass and momentum at
each node can be expressed as follows
n :
6
i=1
i
(n) = 0 (3.2.17)
and
n :
6
i=1
c
i
i
(n) = 0. (3.2.18)
These equations imply corresponding conservation laws for the Boolean ar-
rays n
i

i
n
i
(t + 1, r +c
i
) =
i
n
i
(t, r) (3.2.19)
and

i
c
i
n
i
(t + 1, r +c
i
) =
i
c
i
n
i
(t, r). (3.2.20)
Exercise 3.2.5. (*)
Derive Eq. (3.2.17).
3.2.3 The Liouville equation
Now lattice-gas cellular automata will be considered from the viewpoint of
statistical mechanics. In classical statistical mechanics the deterministic de-
scription of systems with many degrees of freedom by Hamiltonian equations
is abandoned and replaced by a probabilistic approach (Gibbs ensemble:
instead of deriving equilibrium values from the long time limit of a single
system, one calculates them as mean values over a large hypothetical set
of equivalent systems
15
; some authors prefer to avoid the introduction of
ensembles: see, for example, Ma, 1993). One can proceed quite similar for
lattice-gas cellular automata. The microscopic description of the FHP model
encompasses already probabilistic elements (choice of the sense of rotation
for 2-particle collisions). To avoid confusion because of these two dierent
types of probabilities the reader should - at least for a while - consider the
microdynamics as a deterministic process (like it is indeed for HPP).
Consider a lattice L of nal extend with periodic (cyclic) boundary condi-
tions. The phase space (Gibbs) is dened as the set of all possible states
s(.) of the lattice L. At time t = 0 an ensemble of initial states is given with
probabilities P (0, s(.)) 0 which add up to 1:
s(.)
P (0, s(.)) = 1.
Each member of the ensemble evolves according to the microdynamics of the
lattice gas. This implies the conservation of probabilities
P (t + 1, os(.)) = P
_
t, c
1
s(.)
_
(3.2.21)
or
P (t + 1, cs(.)) = P (t, s(.)) . (3.2.22)
Eq. (3.2.22) is called the Liouville equation because of its close analogy to
the Liouville equation in classical statistical mechanics (compare page 139).
In the case of nondeterministic microdynamics like for FHP the Liouville
equation has to be replaced by the more general Chapman-Kolmogorov equa-
tion
P (t + 1, os(.)) =
s(.)
rL
A(s(r) s
(r)) P (t, s(.)) . (3.2.23)

15
Gibbs (1902) writes in his preface: For some purposes, however, it is desirable to
take a broader view of the subject. We may imagine a great number of systems
of the same nature, but diering in the congurations and velocities which they
have at a given instant, and diering not merely innitesimally, but it may be
so as to embrace every conceivable combination of conguration and velocities.
And here we may set the problem, not to follow a particular system through its
succession of congurations, but to determine how the whole number of systems
will be distributed among the various conceivable congurations and velocities
at any required time, when the distribution has been given for some one time.
Exercise 3.2.6. (*)
How many dierent states are possible a) at a single node and b) on a lattice
with N nodes? When does this number exceed 10
10
?
3.2.4 Mass and momentum density
In the framework of the probabilistic description ensemble mean values
q (n(t, ...)) for observables q are dened by
q (n(t, ...)) :=
s(.)
q (s(.)) P (t, s(.))
By far the most important observables are the mean occupation numbers
N
i
(t, r) = n
i
(t, r)
which are used to dene the mass
(t, r) :=
i
N
i
(t, r)
and momentum density
j(t, r) :=
i
c
i
N
i
(t, r).
These quantities are dened with respect to nodes and not to cells or area
16
.
The density per cell d is calculated by division of by the number of cells
per node b (= 6 for FHP-I):
d =

b
. (3.2.24)
The ow velocity is dened by the (non-relativistic) relation momentum den-
sity = mass density velocity:
j(t, r) = (t, r) u(t, r).
Of course the microscopic conservation equations (3.2.19) and (3.2.20) imply
conservation of the averaged quantities
i
N
i
(t + 1, r +c
i
) =
i
N
i
(t, r), (3.2.25)
i
c
i
N
i
(t + 1, r +c
i
) =
i
c
i
N
i
(t, r) (3.2.26)
16
The area per node is

3/2 (compare Fig. 3.2.1).
3.2.5 Equilibrium mean occupation numbers
After many denitions in the previous subsections now one of the main results
of the theoretical analysis of the FHP model will be derived, namely the
equilibrium occupation numbers N
eq
i
. Frisch et al. (1987) proved the following
theorem which is valid for HPP, FHP and FCHC:
Theorem 3.2.1. (Frisch et al., 1987)
The following statements are equivalent:
1. The N
eq
i
s are a solution of equation (3.2.23).
2. The N
eq
i
s are a solution of the set of b equations
i = 1, ..., b :
i
(N) :=
ss
(s
i
s
i
) A(s s
j
N
s
j
j
(1 N
j
)
(1s
j
)
= 0.
(3.2.27)
3. The N
eq
i
s are given by the Fermi-Dirac distribution
N
eq
i
=
1
1 + exp (h +q c
i
)
(3.2.28)
where h is a real number and q is a D-dimensional vector.
Proof. The complete proof is given in Appendix C of Frisch et al. (1987).
Here only the step from 2. to 3. will be discussed. The semi-detailed balance
condition and the nonexistence of spurious invariants has to be taken into
account. In the following the superscript eq will be dropped in order to keep
the notation simple.
Dene
N
i
:=
N
i
1 N
i
and
:=
b
j=1
(1 N
j
).
Eq. (3.2.27) may be written as
ss
(s
i
s
i
)A(s s
N
s
j
j
= 0. (3.2.29)
Multiply Eq. (3.2.29) by

N
i
, sum over i and use
i
(s
i
s
i
) log

N
i
= log
N
s
j
j
N
s
j
j
to obtain
ss
A(s s
) log
_
_
N
s
j
j
N
s
j
j
_
_
N
s
j
j
= 0. (3.2.30)
Semi-detailed balance
s
A(s s
) =
A(s s
) = 1
implies that
ss
A(s s
)
_
_
N
s
j
j

N
s
j
j
_
_
= 0. (3.2.31)
Combining Eqs. (3.2.30) and (3.2.31), one obtains
ss
A(s s
)
_
_
log
_
_
N
s
j
j
N
s
j
j
_
_
N
s
j
j
+
N
s
j
j

N
s
j
j
_
_
= 0. (3.2.32)
The relation (x > 0, y > 0)
y log
x
y
+y x =
_
y
x
log
t
y
dt 0 (3.2.33)
where equality being achieved only when x = y will be exploited. The left
hand side of (3.2.32) is a linear combination of expressions of the form (3.2.33)
with nonnegative weights A(s s
). For it to vanish, one must have
N
s
j
j
=
N
s
j
j
whenever A(s s
) ,= 0.
This is equivalent to
s, s
i
log(

N
i
)(s
i
s
i
)A(s s
) = 0. (3.2.34)
Eq. (3.2.34) means that log(

N
i
) is a collision invariant. Now assuming that
only mass and momentum are conserved and no spurious invariants exist,
one concludes that
log(

N
i
) = (h +q c
i
),
which is the most general collision invariant (a linear combination of the
mass invariant and of the D momentum invariants). Reverting to the mean
populations N
i
=

N
i
/(1 +

N
i
), one obtains (3.2.28). q.e.d.
Calculation of the Lagrange multipliers at small ow speeds. The
determination of the Lagrange multipliers
17
h and q is constrained by the
conserved quantities
=
i
N
i
=
i
1
1 + exp (h +q c
i
)
(3.2.35)
and
u =
i
N
i
c
i
=
i
c
i
1 + exp (h +q c
i
)
. (3.2.36)
These equations apply also for HPP where the problem of the determination
of the Lagrange multipliers can be reduced to a cubic equation (Hardy et al.,
1973). In contrast to HPP, for the FHP model explicit solutions are known
only in a few special cases.
The Lagrange multipliers can be calculated by an expansion for small Mach
numbers M
a
:= U/c
s
, i.e. for speeds U = [u[ well below the sound speed c
s
.
The somewhat lengthy calculations in Appendix 6.2 apply to FHP and HPP.
The equilibrium distributions for FHP-I read
N
eq
i
(, u) =

6
+

3
c
i
u +G()Q
i
u
+O
_
u
3
_
(3.2.37)
with
G() =
1
3
6 2
6
(3.2.38)
and
Q
i
= c
i
c
i

1
2
. (3.2.39)
The term quadratic in u will lead to the nonlinear advection term in the
Navier-Stokes equation. This is the reason to expand the N
eq
i
up to second
order.
Frisch et al. (1987) have shown that the generalization of (3.2.37) to (3.2.39),
namely
N
eq
i
(, u) =

b
+
D
c
2
b
c
i
u
i
+G()Q
i
u
+O
_
u
3
_
(3.2.40)
with
G() =
D
2
2c
4
b
b 2
b
(3.2.41)
and
Q
i
= c
i
c
i

c
2
D
. (3.2.42)
(b number of cells per node, D dimension) is valid for HPP, FHP-I, FHP-II,
FHP-III and FCHC.
17
The Fermi-Dirac distribution (3.2.28) could be derived from a maximum entropy
principle whereby the constraints of mass and momentum conservation are cou-
pled to the entropy (Shannon) by Lagrange multipliers like h and q. This method
will be discussed in some detail in Section 4.3 and applied in Section 5.2.
Exercise 3.2.7. (*)
Specify the equilibrium distribution for HPP and compare it with the linear
distribution (3.1.13) which was used for initialization in Section 3.1.
3.2.6 Derivation of the macroscopic equations: multi-scale analysis
The derivation of macroscopic equations by multi-scale analysis is one of
the most demanding topics in the whole book. In the current section the
expansion will be followed up to rst order only. The calculation of all second
order terms is too involved to be shown here in detail. The multi-scale analysis
will be discussed again in a special section devoted to the Chapman-Enskog
expansion (Section 4.2) and will be applied up to second order in Section 5.2
in the context of lattice Boltzmann models.
In the former subsection the distribution functions for a global (homogeneous)
equilibrium were derived. The interesting aspects of uid ows and of nature
in general lie, however, in its variations in space and time. One can think of
the real world as a patchwork of thermodynamic equilibria whose parameters
like mass, momentum or energy
18
density show slow changes in space, such
that every point can be characterized by the local values of mass, momentum
and energy density.
For FHP equilibrium mean occupation numbers N
eq
i
have been derived which
depend continuously on tunable parameters, namely the mean values of the
conserved quantities mass and momentum density. At the beginning of a
numerical simulation a distribution of mass and momentum density will be
initialized which varies on large (compared to the lattice unit
19
) spatial scales
1
(measured in lattice units; is a small number). In the course of time
three phenomena can be distinguished with respect to their characteristic
time scales:
1. Relaxation toward local equilibrium with time scale
0
: very fast. Few
collision are necessary to reach local equilibrium (compare Fig. 3.2.6).
Note that the number of collisions per time interval is a function of mass
density. Thus the characteristic time scale
0
is large at low and high
mass densities where collisions are rare.
2. Sound waves (perturbations of mass density) and advection with time
scale
1
: fast, but slower than relaxation toward local equilibrium.
3. Diusion with time scale
2
: distinctly slower than sound waves and
advection.
18
Energy does not play a role for FHP and many other LGCA because it does
not exist as an independent quantity or because it is not conserved by certain
collisions whereas so-called thermal LGCA include an energy equation (compare
Section 3.7).
19
Lattice unit = distance between neighboring nodes = 1 for FHP
Let us consider, for example, spatial variations on the scale of
1
= 100 lat-
tice units which may be due to an obstacle in a ow (compare the von Karman
vortex street shown in Fig. 3.6.5). The relaxation toward local equilibrium
proceeds on a time-scale of order
0
= 1, i.e. in a few time steps
20
. Sound
waves
21
and advection show time scales of the order of
1
= 100 whereas
diusion with time-scales of
2
= 10 000 is much slower.
The microdynamics of the lattice-gas cellular automata contain all of these
phenomena. We will now apply the so-called multi-scale technique to pick out
the processes of interest here, namely the hydrodynamic modes. Correspond-
ing to the dierentiation given above one introduces three time variables:
t
(discrete)
t
1
= t
t
2
=
2
t
where the last two scales will be considered as continuous

22
variables. With
respect to space only two scales have to be distinguished because sound waves
and advection as well as diusion act on similar spatial scales:
r
(discrete)
r
1
= r
(continuous)
N
(0)
i
will refer to mean occupation numbers for given local values of and
u. The N
(0)
i
are given by the global form of the equilibrium distributions
(3.2.37) but where and u are the local values of mass density and velocity,
i.e. N
(0)
i
(r, t) N
eq
i
(, u). The actual occupation numbers N
i
(t, r) are close
to the equilibrium values and therefore can be expanded about N
(0)
i
:
N
i
= N
(0)
i
(t, r) +N
(1)
i
(t, r) +O
_
2
_
(3.2.43)
Terms of higher than linear order in will be neglected. The linear corrections
do not contribute to the local values of mass and momentum density:
i
N
(1)
i
(t, r) = 0 and
i
c
i
N
(1)
i
(t, r) = 0 (3.2.44)
which follows from
20
At reasonable mass densities, because otherwise there are not enough collisions.
21
Impressive sound waves are created, for example, when an obstacle like the plate
in Fig. 3.6.5 is suddenly put into an initially homogeneous ow.
22
Over longer time scales a thing or two are smoothed out.
=
i
N
i
(t, r) =
i
N
(0)
i
(t, r) and (3.2.45)
j =
i
c
i
N
i
(t, r) =
i
c
i
N
(0)
i
(t, r). (3.2.46)
In the Chapman-Enskog expansion of the Boltzmann equation (Chapman,
1916, 1918; Enskog, 1917; Chapman and Cowling, 1970; Cercignani, 1990)
is the Knudsen number K
n
, i.e. the ratio between the mean free path length
l and the characteristic length scale of the system L which can be the diam-
eter of an obstacle (for example: ow past a sphere) or the size of the whole
domain. The hydrodynamic (continuous) regime is characterized by small
Knudsen numbers whereas nite size eects play a role at Knudsen number
of order 1 or higher (Knudsen ows).
The microscopic conservation laws (3.2.25) and (3.2.26) are the starting point
for the multi-scale analysis. The mean populations after collision and prop-
agation are expanded up to second order in around its values before the
collision step:
_
1
c
i
_
N
i
(t + 1, r +c
i
) =
_
1
c
i
_
[ N
i
(t, r) +
t
N
i
+c
i
N
i
+
1
2
t
N
i
+
1
2
c
i
c
i
N
i
(3.2.47)
+c
i
N
i
+O
_
3
N
i
_
]
In what follows the fast (local) relaxation processes will be neglected in the
theoretical description because one is only interested in the hydrodynamic
behavior of the lattice-gas cellular automata. The derivations in time and
space are substituted in Eq. (3.2.47) according to the scalings given above:
t

(1)
t
+
2
(2)
t
(3.2.48)

(1)
x
. (3.2.49)
Insertion of Eqs. (3.2.43), (3.2.47), (3.2.48) and (3.2.49) into Eqs. (3.2.25)
and (3.2.26) leads to
i
_
1
c
i
_
[N
i
(t + 1, r +c
i
) N
i
(t, r)]
=
i
_
1
c
i
_
_
(1)
t
N
(0)
i
+
2
(1)
t
N
(1)
i
+
2
(2)
t
N
(0)
i
+
3
(2)
t
N
(1)
i
+c
i
(1)
x
N
(0)
i
+
2
c
i
(1)
x
N
(1)
i
+
1
2
(1)
t

(1)
t
N
(0)
i
+
1
2
(1)
t

(1)
t
N
(1)
i
+
3
(1)
t

(2)
t
N
(0)
i
+
4
(1)
t

(2)
t
N
(1)
i
+
4
(2)
t

(2)
t
N
(0)
i
+
5
(2)
t

(2)
t
N
(1)
i
+
1
2
2
c
i
c
i
(1)
x
(1)
x
N
(0)
i
+
1
2
3
c
i
c
i
(1)
x
(1)
x
N
(1)
i
+
2
c
i
(1)
t

(1)
x
N
(0)
i
+
3
c
i
(1)
t

(1)
x
N
(1)
i
+
3
c
i
(2)
t

(1)
x
N
(0)
i
+
4
c
i
(2)
t

(1)
x
N
(1)
i
_
= 0.
To rst order in one obtains
(1)
t
i
N
(0)
i
+
(1)
x
i
c
i
N
(0)
i
= 0 (3.2.50)
and
(1)
t
i
c
i
N
(0)
i
+
(1)
x
i
c
i
c
i
N
(0)
i
= 0 (3.2.51)
or
(1)
t
+
(1)
x
(u
) = 0 (continuity equation) (3.2.52)
(1)
t
(u
) +
(1)
x
P
(0)
= 0 (3.2.53)
where
P
(0)
i
c
i
c
i
N
eq
i
=
+ G()T
(MA)
. .
advection term
+O
_
u
4
_
(3.2.54)
is the momentum ux tensor in rst approximation. The momentum advec-
tion tensor
23
T
(MA)
is a tensor of 4th rank
T
(MA)
i
c
i
c
i
Q
i
. (3.2.55)
It is isotropic and given by (compare Section 3.3, Eq. (3.3.10)):
T
(MA)
i
c
i
c
i
Q
i
=
i
c
i
c
i
_
c
i
c
i

1
2
_
=
3
4
(
) .
23
The name is derived from the fact that this tensor is part of the nonlinear
advection term. It occurs, however, also in the dissipative terms of the Navier-
Stokes equation.
Accordingly the components of the momentum ux tensor in rst approxi-
mation read
P
(0)
xx
= G()
_
u
2
v
2
_
+

2
P
(0)
yy
= G()
_
v
2
u
2
_
+

2
(3.2.56)
P
(0)
xy
= P
(0)
yx
= G()2uv
whereas the momentum ux tensor in the real world (i.e. in the Navier-
Stokes equation) is
P
xx
= u
2
+p
P
yy
= v
2
+p (3.2.57)
P
xy
= P
yx
= uv.
Identication of

2
(1 g()u
2
) with the pressure p leads to
P
(0)
xx
= g()u
2
+p
P
(0)
yy
= g()v
2
+p (3.2.58)
P
(0)
xy
= P
(0)
yx
= g()uv
which looks similar to the momentum ux tensor (3.2.57) except for the
factor g() = G()/2. For small values of u
2
the pressure is given by the
isothermal relation
p =

2
= c
2
s
(3.2.59)
with the sound speed c
s
= 1/
2. It can be shown that invariance under

Galilei transformations constrains the g-factor to be equal to 1 (compare
Exercise 3.2.8). Here
g() =
3
6
(3.2.60)
is always smaller than 1 (actually smaller than 1/2). Similar expressions apply
to other lattice models. Thus this g-factor breaks the Galilean invariance.
The deviation of g from 1 is caused by the underlying lattice which is only
invariant under certain discrete translations but not under arbitrary Galilei
transformations. It will be shown later that this disease can be cured when
other distribution functions (Boltzmann instead of Fermi-Dirac) are applied.
This is the case in lattice Boltzmann models. The occurance of Fermi-Dirac
distributions in LGCA is a consequence of the exclusion principle. Therefore
if one sticks to this essential feature of any LGCA the g-disease can be treated
only symptomatically
24
, namely by a rescaling of time
t
t
g()
. (3.2.61)
The equation for the incompressible regime can be derived from the com-
pressible equation by ignoring all density variations except in the pressure
term
25
. Setting to the constant and uniform value
0
and applying the
rescaling of time leads to
0
g(
0
)
u
t
+g(
0
)(u)u =
_
2

0
2
g(
0
)u
2
_
or
u
t
+ (u)u = P (3.2.62)
with the kinematic pressure
P =
_

2
0
g(
0
)
u
2
_
(3.2.63)
Eq. (3.2.62) is the Euler equation (Navier-Stokes without dissipation) of the
FHP model. Eq. (3.2.52) is the continuity equation for the mass density . It
will not change when terms of order
2
are included.
The terms of order
2
are calculated by Frisch et al. (1987) and Henon
(1987b). Adding up terms of order and
2
while neglecting terms of or-
der u
3
,
2
u
2
and
3
u leads in the same incompressible limit as discussed
before to the Navier-Stokes equation
u = 0
t
u + (u)u = P +
2
u
(3.2.64)
where is the (scaled) kinematic shear viscosity ( =
(u)
/g(
0
); the un-
scaled shear viscosity
(u)
for dierent FHP models is given in Table 3.2.2;
compare also Fig. 3.2.4). Thus we have attained the object of our desire: the
incompressible Navier-Stokes equation!
Please note that the FHP models do not possess an energy equation. In the
FHP-I model mass and (kinetic) energy conservation are essentially identical.
In models with rest particles, some collisions do not respect conservation of
Fig. 3.2.3. The Reynolds coecient R
as a function of the density per cell d for

FHP-I, FHP-II and FHP-III. FHP-I: 2-particle and symmetric 3-particle collisions;
FHP-II: additional collisions involving rest particles and 2-particle collisions with
spectator; FHP-III: additional 4-particle collisions (collision saturated model). The
Reynolds coecient is dened as R
:=
R
e
L M
a
=
c
s
g(d)
(d)
; R
max
= max
d
(R
) =
R
(d
max
); d is the density per cell (Eq. 3.2.24), c
s
sound speed, g(d) density de-
pendent g-factor, and
(u)
(unscaled) kinematic shear viscosity.
0 0.1 0.2 0.3 0.4 0.5
0
0.5
1
1.5
2
2.5
Density per cell
R
e
y
n
o
l
d
s

c
o
e
f
f
i
c
i
e
n
t
FHPI
FHPII
FHPIII
(kinetic) energy. lattice-gas cellular automata with energy equations will be
discussed in Section 3.7.
24
The symptomatic treatment of the g-disease does not solve all problems.
DHumières et al. (1987) have shown that vorticity is advected at speed g(d)u =
u. In order to x this problem they have proposed a model with 8 bits over the
FHP lattice (see Subsection 3.2.10 for details).
25
For some ne points see Majda, 1984.
Table 3.2.2. Analytical values for three dierent FHP models.
(u)
(unscaled)
kinematic shear viscosity;
(u)
(unscaled) kinematic bulk (compressional) viscosity
(taken from Frisch et al., 1987). d is the density per cell (Eq. 3.2.24).
FHP-I FHP-II FHP-III
Number of cells 6 7 7
c
s
1/
3/7
3/7
g(d)
1
2
1 2d
1 d
7
12
1 2d
1 d
7
12
1 2d
1 d
(u)
1
12
1
d(1 d)
3

1
8
1
28
1
d(1 d)
3
1
1 4d/7

1
8
1
28
1
d(1 d)
1
1 8d(1 d)/7

1
8
(u)
0
1
98
1
d(1 d)
4

1
28
1
98
1
d(1 d)
1
1 2d(1 d)

1
28
R
max
0.387 1.08 2.22

d
max
0.187 0.179 0.285
Fig. 3.2.4. The g-factor, unscaled kinematic bulk viscosity (
(u)
), and the unscaled
(
(u)
) and scaled () kinematic shear viscosity as functions of the density per cell
(d) for the models FHP-I, FHP-II and FHP-III (compare Table 3.2.2). The bulk
viscosity vanishes in FHP-I.
0 0.2 0.4
0
0.2
0.4
0.6
0.8
Density per cell d
g
(
d
)

f
a
c
t
o
r
FHPI
FHPII and FHPIII
0 0.2 0.4
0
0.1
0.2
0.3
0.4
Density per cell d
B
u
l
k

v
i
s
c
o
s
i
t
y

(
u
)
FHPII
FHPIII
0 0.2 0.4
0
0.5
1
1.5
2
Density per cell d
S
h
e
a
r

v
i
s
c
o
s
i
t
y

(
u
)
FHPI
FHPII
FHPIII
0 0.2 0.4
0
2
4
6
Density per cell d
S
h
e
a
r

v
i
s
c
o
s
i
t
y

FHPI
FHPII
FHPIII
Show that the generalized (g-factor) substantial derivative
u
t
+g (u) u (3.2.65)
is invariant under Galilei transformations if and only if g 1.
Hint: it will be sucient to consider the special Galilei transformation
x
= x +ct; c = const
y
= y
(3.2.66)
z
= z
t
= t
Calculate the components of the momentum ux tensor P
(0)
for HPP from

the formula
G()T
(MA)
+
c
2
D
. (3.2.67)
Whats wrong with the advection term?
3.2.7 Boundary conditions
The coding of boundary conditions (BC) is an essential part of the LGCA
(and any other numerical) method. There are at least ve dierent types of
BC:
1. Periodic BC are often used even if it is not realistic because they are so
easy to code.
2. Inow BC (example: channel ow).
3. Outow BC (example: channel ow) can be very dicult to deal with,
especially when waves try to leave the model domain (compare, for ex-
ample, Orlanski, 1976, Red and Smedstad, 1984, and Stevens, 1991).
4. No-slip BC (u = 0) apply to solid boundaries (walls, obstacles).
5. Slip BC, i.e. the velocity component normal to the boundary and the
normal derivative of the tangential component vanish (u
n
= 0 and
u
t
/n = 0), apply to solid boundaries where the frictional force ad-
jacent to the wall is not resolved.
Even when constraints are formulated only for the velocity one usually also
requires conservation of mass
26
. Whereas in channel ows a small violation
of mass conservation could be tolerable, because each uid element leaves the
domain after some time anyway, such a violation is not acceptable for ows
in closed domains where a small but steady leakage, for example, would lead
to an empty basin after a while.
The coding of the rst two types of BC is obvious not only for FHP but for
any kind of LGCA or LBM and therefore needs no further comment. Since
the beginning of simulations with FHP the following heuristic procedures for
the implementation of no-slip and slip conditions are used (dHumières and
Lallemande, 1987):
No-slip: Collision on boundary points are skipped. Instead the incoming
particles are turned around by 180
. In the next propagation step they will

26
However, momentum is most often not conserved at the boundaries.
leave the node in their former incoming direction. The mean value over in-
state and out-state yields u = 0. This ipping of the incoming particles is
also called bounce-back rule. The kinetic equation for this implementation
of the no-slip condition reads
n
i
(x
b
+c
i
, t + 1) = n
i+3
(x
b
, t)
where x
b
is a boundary point and as usual the indices are understood mod-
ulo 6.
This node-oriented implementation of the no-slip BC has the advantage
that it is independent of the orientation of the wall (for alternative imple-
mentations see Rem and Somers, 1989, Cornubert et al., 1991, and Ziegler,
1993).
Slip: Collisions on boundary points are skipped. Instead the incoming par-
ticles are reected like a light ray (specular reection) at the wall. The
mean value over in-state and out-state yields u
n
= 0 and the tangential
momentum of the particles is conserved. The kinetic equation for this im-
plementation of the no-slip condition reads
n
i
(x
b
+c
i
, t + 1) = n
i3
(x
b
, t).
Only several years later Cornubert et al. (1991) have investigated discrepan-
cies of the distribution functions in the boundary layer due to these kind of im-
plementations of boundary conditions. For FHP they found anisotropic Knud-
sen layers adjacent to obstacles. The eective boundary of obstacles is not
identical with the node locations but lies somewhat outside the outer nodes
of the obstacle. The precise location depends on the direction (anisotropic)
and is smaller than the distance to the next neighbor node in the uid region.
Especially for small obstacles this eect has to be taken into account.
3.2.8 Inclusion of body forces
In principle body forces, i.e. a change of momentum, can be applied on the
macroscopic or microscopic level. The macroscopic method, however, requires
at each time step the calculation of mean values (coarse graining), change of
the momentum j and re-initilization. Of course this procedure is computa-
tionally much too demanding. Thus only a microscopic method is feasable.
Body forces F which may vary in space and time but are independent of the
ow velocity u can be realized by ipping particles with velocity c
i
into
particles with velocity c
i
(for forces parallel to c
i
). The probability of this
ipping has to be proportional to the magnitude of F.
The following results are from Appendix D in Frisch et al. (1987). Boolean
transition variables
ss
are dened such that their mean values
ss
= B(s s
) (3.2.68)
are a set of transition probabilities associated to the body-force. The transi-
tion probabilities satisfy normalization
B(s s
) = 1 (3.2.69)
and mass conservation
i
(s
i
s
i
) B(s s
) = 0 s, s
, (3.2.70)
but do not satisfy momentum conservation (which is desired of course!), semi-
detailed balance and (-invariance. In case I (body-force f independent of
velocity) they are further constrained by
f =
s,s
,i
c
i
(s
i
s
i
) B(s s
)
_
d
1 d
_
p
(1 d)
b
(3.2.71)
where p =

j
s
j
and b is the number of cells per node. If f is space and/or
time dependent, so are the B(s s
)s.
In case of a force linear in velocity
f
= C
(3.2.72)
the B(s s
)s are constrained by
0 =
s,s
,i
c
i
(s
i
s
i
) B(s s
)
_
d
1 d
_
p
(1 d)
b
, p =
j
s
j
, (3.2.73)
and
C
=
D
c
2
(1 d)
b1

s,s
,i
c
i
(s
i
s
i
) B(s s
)
_
d
1 d
_
p
j
s
j
c
j
(3.2.74)
where p =

j
s
j
.
In order to illustrate the method let us consider the simplest example, namely
time-independent homogeneous forcing in the direction of a particular lattice
velocity, say in x-direction. We can ip particles with velocity c
6
= c
3
=
(1, 0) into particles with velocity c
6
= (1, 0) whenever this is possible while
leaving all other particles unchanged (compare Fig. 3.2.5). Already after one
time step the domain mean x-velocity, u
x
(t), increased from zero (its initial
value) to approximately 0.2. This extreme acceleration leads to high Mach
numbers after a few time steps. Thus in simulations of incompressible ow
problems the ipping rate has to be lowered.
Fig. 3.2.5. Inclusion of body-forces: The plot shows the increase of domain mean
x-velocity, u
x
(t), due to ipping of particles with velocity c
6
= c
3
= (1, 0) into
particles with velocity c
6
= (1, 0) whenever this is possible (FHP-I, d = 0.3).
0 20 40 60 80 100 120
0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Time t
u
x

(
t
)

(
m
e
a
n

o
v
e
r

d
o
m
a
i
n
)
3.2.9 Numerical experiments with FHP
Here the results of some numerical calculations will be discussed. The code
written in C is accessible (see the web address given in the Preface).
1. Relaxation toward equilibrium (compare program exper1.c): The FHP-
I model is initialized with a distribution which is spatially homogenous
but far from equilibrium. At a density 2 enough collisions occur in
order to drive the occupation numbers toward their equilibrium values
(compare Fig. 3.2.6).
Fig. 3.2.6. Relaxation toward equilibrium for FHP-I: The model is initialized with
a distribution which is spatially homogenous but far from equilibrium. The domain
size is 320 times 320 nodes. For simplicity periodic boundary conditions are applied.
At a density 2 enough collisions occur in order to drive the occupation numbers
toward their equilibrium values (dashed lines) which are almost reached after 10
time steps. The mean values over the time levels 30 to 60 compares very well with
the theoretical values according to equilibrium distribution (3.2.37).
0 10 20 30 40 50 60
0.28
0.3
0.32
0.34
0.36
0.38
0.4
N
1
(3060)
(*) = 0.37482 N
1
(eq)
= 0.37593
N
2
(3060)
(o) = 0.34111 N
2
(eq)
= 0.34167
N
3
(3060)
(+) = 0.30021 N
3
(eq)
= 0.30069
N
4
(3060)
(x) = 0.29126 N
4
(eq)
= 0.29244
N
5
(3060)
(s) = 0.3237 N
5
(eq)
= 0.32418
N
6
(3060)
(d) = 0.36613 N
6
(eq)
= 0.36669
Time steps

M
e
a
n

o
c
c
u
p
a
t
i
o
n

n
u
m
b
e
r
s

N
i
FHPI;
m
= 1.9972; u
m
= 0.049566; v
m
= 0.043785
2. Propagation of sound waves: Here sound waves are excited by a density
perturbation (compare Fig. 3.2.7 upper left). The waves propage isotrop-
ically with a speed of c
s
= 1/
2.
Fig. 3.2.7. Propagation of sound waves: The FHP-I model is initialized with van-
ishing velocity and constant density ( 2) except for a positive radial symmetric
density perturbation near the center of a domain with 960 times 960 nodes. The
plots show the mean densities calculated over subdomains (macrocells) of 32 times
32 nodes at four dierent time levels. The values of the contour lines are always
2.1, 2.2, 2.3, 2.4, 2.5. The propagation of the density perturbation is isotropic and
the speed of propagation is consistent with the sound speed c
s
= 1/
2 (in 300 time

steps the sound will propagate over a distance of 300/32/
2 13 sidelengths of
macrocells).
5 10 15 20 25 30
5
10
15
20
25
30
Initial mass density
max
= 2.6709
5 10 15 20 25 30
5
10
15
20
25
30
max
= 2.667; t = 100
5 10 15 20 25 30
5
10
15
20
25
30
max
= 2.2725; t = 200
5 10 15 20 25 30
5
10
15
20
25
30
max
= 2.2559; t = 300
3. Flow past an obstacle: Our next goal is the simulation of the ow past
an obstacle. Wether the ow is laminar or turbulent depends on the
Reynolds number R
e
which is dened by
R
e
:=
U L
(3.2.75)
where U is a characteristic ow speed (usually the ow speed far upstream
of the obstacle), L is a characteristic spatial scale of the obstacle (for an
obstacle in the form of a circular cylinder, for example, L is the radius
or diameter of the cross-section), and is the kinematic shear viscosity.
From experiments it is known that eddies are formed and shedded when
the Reynolds number becomes larger than about 50. A so-called von
Karman vortex street will build up (compare the beautiful pictures in
the book of van Dyke, 1982).
How can one simulate a ow with R
e
90 with FHP-II? The ow speed
has to be small compared to the sound speed c
(FHP-II)
s
=
_
3/7 0.65.
So u = 0.2 is a good value. The density per cell d should be below 0.5
(the scaling factor g(d) vanishes at d = 0.5) and is chosen here as d = 2/7
which results in a density per node = 2. The scaled kinematic viscosity
at this density is approximately 0.8. The only free parameter is the size
of the obstacle L which can be calculated from
L =
R
e

u
=
88 0.88
0.2
= 388 (3.2.76)
(see results in Fig. (3.2.8)).
Fig. 3.2.8. Flow past a plate. The FHP-II model is initialized with a homogenous
ow in x-direction with speed u = 0.2 (in lattice units). The lattice encompasses
4096 times 1792 nodes. The width of the plate is 388. At a density per node = 2.1
the (scaled) kinematic viscosity is 0.88. Thus the Reynolds number is 88. The gure
shows the ow minus the mean ow velocity after t = 260, 000 time steps.
Please note that the type of boundary conditions are essential. If one
applies periodic boundary conditions the simulation addresses ow past
a periodic array of cylinders. In contrast to ow past a single cylinder
the ow may be steady even at a Reynolds number of 100 (Gallivan et
al., 1997).
Fig. 3.2.9. Flow past a circular cylinder. Lattice with 6400 times 2560 nodes. The
uppermost plot shows the result after 20000 time steps for the FHP-I model (the
ow velocity averaged over the whole domain has been substracted in both plots in
order to make the pertubations better visible). If 3-particle collisions are left out the
ow eld looks quite dierent (lower plot).
Exercise 3.2.10. (*)
Consider typical ow velocities in the atmosphere and oceans. At which spa-
tial scales the Reynolds numbers are 1, 1000 and 10
6
?
3.2.10 The 8-bit FHP model
DHumières et al. (1987) proposed a model with two populations (n
0
, m
0
)
that represent rest particles. With 2 bits up to 3 rest particles can be de-
scribed: (n
0
, m
0
) = (0,0) no, (1,0) one, (0,1) two and (1,1) three rest par-
ticles. All collisions conserving mass and momentum are included (collision
saturated). The collisions a) to c) in Table 3.2.2 proceed independent of the
number of rest particles. The collisions leading to creation and destruction
of rest particles are all included, except a few cases which take place with
probability x, y, or z (compare Tables 3.2.10 and 3.2.11). It can be shown
that for x = 0.5, y = z = 0.2, there exists a value for d for which g(d) = 1
and dg(d)/dd = 0. DHumières et al. recommend a model with x = 1/2,
y = z = 1/6 which gives g = 1.0 at d = 0.21. The model violates semi-detailed
balance (compare Exercise 3.2.11). Numerical experiments show that vortic-
ity is advected close to the ow speed U
0
. The 8-bit-FHP model is probably
already too complex to allow coding with bit-operators. The use of look-up
tables makes it much more clumpsy than the 7-bit-FHP models.
Exercise 3.2.11. (*)
Show that the 8-bit FHP model of dHumières et al. (1987) violates semi-
detailed balance.
Exercise 3.2.12. (***)
Repeat the simulation of the ow past a cylinder with the FHP-II model.
Compare the computational time for constant Reynolds number by varying
the upstream velocity or the size of the system. Analyze the dierences in
the ow pattern when applying the random or the chiral two-particle head-on
collision.
Fig. 3.2.10. FHP-8-bit model: collisions involving rest particles; occupied cells are
represented by arrows, empty cells by thin lines, the number of rest particles is
indicated by the number in the central circle.
f) _
1
_
e
e
e
e
x
1
_
_
_
0
/
/
`
`*
e
e
g) _
2
_
e
e
e
e
y
1
_
_
_
1
/
/
`
`*
e
e
h) _
3
_
e
e
e
e
z
1
_
_
_
2
/
/
`
`*
e
e
i) j
3
/
/
`
`*
e
e
z/2
.
.
.
-
-
-
_
2
/
/
/
/
_
e
e
e
e
_
2
_
`
`
`
`*

Fig. 3.2.11. FHP-8-bit model: collisions involving rest particles (continued); occu-
pied cells are represented by arrows, empty cells by thin lines, the number of rest
particles is indicated by the number in the central circle.
j) j
r
/
/
/
/
_
e
e
e
e
1
_ r = 0, 1 or 2
s = r + 1
_
s
/
/
`
`*
e
e
k) j
3
/
/
/
/
_
e
e
e
e
1-z
z
.
.
.
-
-
-
_
3
/
/
/
/
_
e
e
e
e
_
2
/
/
`
`*
_ _
e
e
l) j
r
/
/
`
`*
_ _
e
e
1/2
1/2
.
.
.
-
-
-
r = 0, 1 or 2
s = r + 1
_
s
_
`
`
`
`*
_
s
_
/
/
/
/
e
e
e
e
m) j
3
/
/
`
`*
_ _
e
e
1
_ _
2
_
/
/
`
`
/
/
`
`*
n) j
r
_
/
/
`
`
/
/
`
`*
1
_ r = 0, 1 or 2
s = r + 1
_
s
/
/
`
`*
_ _
e
e

3.3 Lattice tensors and isotropy in the macroscopic
limit
All lattice-gas cellular automata are based on extremely discretized phase
spaces. Nevertheless it is to be expected that the spatial discretization will
be smoothed out on scales which are much larger than the grid spacing
27
.
Things are less obvious for the discretization of the velocities. Calculation of
mean values over an angular direction is restricted to a range of 2 which
contains only a few velocities (six for FHP). The multi-spin coding yields
tensors which consist of the components of the lattice velocities. These ten-
sors are invariant with respect to elements of the associated nite symmetry
group but in general not with respect to arbitrary orthogonal transforma-
tions (including continuous rotations). A sucient condition for reasonable
macroscopic equations encloses the isotropy of lattice tensors (to be dened
below) of 2nd and 4th rank. The lattice tensors with odd rank vanish because
of the symmetry of the lattices.
Group theoretical methods allow quite general propositions concerning the
isotropy of tensors of this special form (Wolfram, 1986). Discussion of group
theoretical concepts is outside the scope of this work. Explicit expressions
for the most general isotropic tensors will be given and lattice tensors and
generalized lattice tensors for the various lattice-gas cellular automata will be
calculated. Note that the results apply also to the lattice Boltzmann models
discussed later on in Section 5.
3.3.1 Isotropic tensors
Denition: A tensor T
2
...
n
of nth rank is called isotropic if it is invari-
ant with respect to arbitrary orthogonal transformations O (rotations and
reections)
T
2
...
n
= T
2
...
n
O
1
O
2
...O
n
. (3.3.1)
The most general isotropic tensors up to 4th rank are provided by the fol-
lowing theorem.
Theorem 3.3.1. (Jereys and Jereys, 1956; Jereys, 1965)
1. There are no isotropic tensors of rank 1 (vectors).
2. An isotropic tensor of rank 2 is proportional to
.
3. An isotropic tensor of rank 3 is proportional to
28
.
27
There is still the problem of a selected reference system which violates Galilean
invariance and which can produce problems in the macroscopic limit.
28
Levy-Civita symbol
:
123
=
231
=
312
= 1,
132
=
321
=
213
= 1, and
zero otherwise.
3.3 Lattice tensors and isotropy in the macroscopic limit 91
4. There are three dierent (linear independent) tensors of rank 4
,
which can be combined to the most general form
T
= a
+b
+c
, (3.3.2)
where a, b and c are arbitrary constants.
A proof of the theorem can be found in Jereys (1965).
Isotropic tensors of rank n 4 consist only of products of second rank
tensors (for example:
and all tensors that result from cyclic per-

mutation of indices) when n is even or of products of and tensors when n
is odd.
In two dimensions the isotropic tensor of rank 4 (3.3.2) has the following
non-vanishing components:
T
1111
= T
2222
= a +b +c,
T
1122
= T
2211
= a,
(3.3.3)
T
1212
= T
2121
= b,
T
1221
= T
2112
= c.
In particular, the tensor
is non-isotropic (
is 1 if all indices are

equal and 0 otherwise; it is a generalization of the Kronecker symbol
).
The same is true for
.
3.3.2 Lattice tensors: single-speed models
Let us dene the following tensors of rank n
L
2
...
n
=
i
c
i
1
c
i
2
...c
i
n
(3.3.4)
where c
i
are the cartesian components of the lattice velocities c

i
. We will
call these tensors the lattice tensors. Because of their special structure they
are invariant with respect to the symmetry group of the lattice and they are
symmetric in all of their indices. From these symmetries it follows that these
tensors can have a maximal number N of independent components of
N =
_
n +D 1
n
_
=
(n +D 1)!
n!(D 1)!
where D is the (spatial) dimension. Example: A symmetric tensor T of rank
2 in two dimensions has at most N =
_
3
2
_
=
3!
2!1!
= 3 independent compo-
nents:
T =
_
a b
b c
_
.
In lattice-gas cellular automata with one speed (HPP, FHP, FCHC) the mo-
mentum advection tensor (MAT) of 4th rank T
(MA)
(compare Eq. 3.2.55)

occurs in the macroscopic form of the momentum balance. It can be rewritten
in terms of the lattice tensors of rank two and four:
T
(MA)
i
c
i
c
i
Q
i
=
i
_
c
i
c
i
c
i
c
i

1
2
c
i
c
i
_
= L

1
2
L
. (3.3.5)
A sucient condition for the isotropy of T
(MA)
is the isotropy of L
and
L
. T
(MA)
is non-isotropic if L
is isotropic while L
is non-isotropic.
Other combinations do not occur in the models considered. In what follows we
use the notation DkQb of Qian et al. (1992) where k is the spatial dimension
and b is the number of lattice velocities (including c
0
= 0 for rest particles).
Square lattice: HPP (D2Q4). Lattice velocities:
c
i
=
_
cos
2i
4
, sin
2i
4
_
i = 1, 2, 3, 4.
The lattice tensor of rank 2
L
HPP
= 2
_
1 0
0 1
_
= 2
is isotropic.
L
HPP
= 2
(3.3.6)
is non-isotropic. As a consequence the HPP model fails to yield the Navier-
Stokes equations in the macroscopic limit.
Fig. 3.3.1. The lattice velocities of the HPP (D2Q4) lattice.
1
2
3
4
Triangular lattice: FHP (D2Q7). The triangular FHP lattice has hexag-
onal symmetry. Lattice velocities:
c
0
= (0, 0)
c
i
=
_
cos
2i
6
, sin
2i
6
_
i = 1, ..., 6 (3.3.7)
Fig. 3.3.2. The lattice velocities of the FHP (D2Q7) lattice.
1 2
3
4 5
6 0
The lattice tensors of rank 2
L
FHP
= 3
(3.3.8)
and rank 4
L
FHP
=
3
4
(
) (3.3.9)
are isotropic. The momentum advection tensor reads
T
(MA)
=
3
4
(
) . (3.3.10)
Thus, FHP yields the Navier-Stokes equation in the macroscopic limit.
FCHC (D4Q24). The investigation of the lattice tensors of the FCHC
model is left to the reader (Exercise 3.3.1).
The 24 lattice velocities of the FCHC model are given by
(1, 1, 0, 0), (1, 0, 1, 0), (0, 1, 1, 0),
(1, 0, 0, 1), (0, 1, 0, 1), (0, 0, 1, 1).
Calculate the lattice tensors of rank 2 and 4 and show that they are isotropic.
3.3.3 Generalized lattice tensors for multi-speed models
Later on multi-speed
29
lattice-gas cellular automata and multi-speed lattice
Boltzmann models will be discussed. The associated lattice tensors of rank 4
are usually non-isotropic because the symmetry group of the corresponding
lattices is not large enough (note that the situation is dierent for the multi-
speed FHP model where the symmetry of each single-speed sub-lattice is large
enough). But isotropy of 4th rank tensors can be recovered by introducing
weights w
i
for the dierent speeds. These generalized lattice tensors
G
2
...
n
=
i
w
i
c
i
1
c
i
2
...c
i
n
(3.3.11)
occur naturally in the multi-scale analysis of multi-speed models. The weights
correspond to dierent occupation numbers for the dierent speeds in the
global equilibrium with vanishing macroscopic velocities.
The following sub-sections can be skipped in the rst reading and should be
revisited when the appropriate multi-speed model is discussed.
29
Models with several dierent speeds have been encountered before, namely the
FHP variants with rest particles (FHP-II, FHP-III). Particles with vanishing
speed have no inuence on the isotropy of the lattice tensors. Therefore only
models with dierent non-vanishing speeds are called multi-speed models.
D2Q9. Lattice velocities:
c
0
= (0, 0),
c
1,3
, c
2,4
= (1, 0), (0, 1), (3.3.12)
c
5,6,7,8
= (1, 1).
Fig. 3.3.3. The lattice velocities of the D2Q9 lattice.
1
2
3
4
5 6
7 8
0
L
(D2Q9)
= 6
is isotropic. The lattice tensor of rank 4 with the following non-vanishing

components
L
(D2Q9)
1111
= L
(D2Q9)
2222
= 6
L
(D2Q9)
1122
= L
(D2Q9)
2211
= 4
L
(D2Q9)
1212
= L
(D2Q9)
2121
= 4
L
(D2Q9)
1221
= L
(D2Q9)
2112
= 4
is non-isotropic.
Introducing the weights w
i
= 1 for speed 1 (i = 1, ..., 4) and w
i
= 1/4 for
speed

2 (i = 5, ..., 8) leads to isotropic generalized lattice tensors of rank 2
G
(D2Q9)
= 3
(3.3.13)
and rank 4
G
(D2Q9)
. (3.3.14)
D2Q13-WB. Weimar and Boon (1996)
c
0
= (0, 0) rest particle
c
1,2
, c
3,4
= (1, 0), (0, 1) 1-particles
c
5,6,7,8
= (1, 1)

2-particles
c
9,10
, c
11,12
= (2, 0), (0, 2) 2-particles
(3.3.15)
Fig. 3.3.4. The lattice velocities of the D2Q13-WB lattice.
3
4
5
8
9
11
12
1 2
6
7
10 0
L
(D2Q13WB)
= 14
is isotropic whereas the lattice tensor of rank 4

L
(D2Q13WB)
= 4 (
) + 26
is non-isotropic. The weights w

i
= 4 for speed 1, w
i
= 5 for speed

2, and
w
i
= 1 for speed 2 lead to isotropic generalized lattice tensors of rank 2
G
(D2Q13WB)
= 36
(3.3.16)
and rank 4
G
(D2Q13WB)
= 20 (
) . (3.3.17)
D2Q21. The D2Q21 lattice was introduced by Fahner (1991).
c
0
c
1,2
, c
3,4
= (1, 0), (0, 1) 1-particles
c
5,6,7,8
= (1, 1)

2-particles
c
9,10
, c
11,12
= (2, 0), (0, 2) 2-particles
c
13,...,16
, c
17,...,20
= (2, 1), (1, 2)

5-particles
(3.3.18)
L
(D2Q21)
= 34

L
(D2Q21)
= 36 (
) 2

i
= 2 for speed 1 and w
i
= 1 otherwise lead
to isotropic generalized lattice tensors of rank 2
G
(D2Q21)
= 36
(3.3.19)
and rank 4
G
(D2Q21)
= 36 (
) (3.3.20)
D3Q15. Multi-speed lattice Boltzmann models in 3D will be dened in Sec-
tion 5.3. The lattice D3Q15 has the following lattice velocities
c
0
= (0, 0, 0) rest particle
c
1,2
, c
3,4
, c
5,6
= (2, 0, 0), (0, 2, 0) (0, 0, 2) 2-particles
c
7,...,14
= (1, 1, 1)

3-particles
(3.3.21)
1 2
3
4
5
6
7
8
9 10
11
12
13
14
15
16
17 18
19 20
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
x
y
z
L
(D3Q15)
= 16

L
(D3Q15)
= 8 (
) + 16

i
= 2 for speed

3 and w
i
= 1 for speed 2
lead to isotropic generalized lattice tensors of rank 2
G
(D3Q15)
= 24
(3.3.22)
and rank 4
G
(D3Q15)
= 16 (
) . (3.3.23)
D3Q19. The model D3Q19 has the following lattice velocities
c
0
= (0, 0)
c
1,2
, c
3,4
, c
5,6
= (1, 0, 0), (0, 1, 0) (0, 0, 1)
c
7,...,10
, c
11,...,14
, c
15,...,18
= (1, 1, 0), (1, 0, 1), (0, 1, 1).
(3.3.24)
L
(D3Q19)
= 10

L
(D3Q19)
= 4 (
) 2

i
= 2 for speed 1 and w
i
= 1 for speed

2
lead to isotropic generalized lattice tensors of rank 2
G
(D3Q19)
= 12
(3.3.25)
and rank 4
G
(D3Q19)
= 4 (
) . (3.3.26)
3.3.4 Thermal LBMs: D2Q13-FHP (multi-speed FHP model)
For thermal lattice Boltzmann models (Navier-Stokes plus energy equation)
isotropic lattice tensors up to rank 6 are required.
c
i
= (0, 0) i = 0
c
i
=
_
cos
2k
6
, sin
2k
6
_
i = 1, 2, ..., 6; k = i (3.3.27)
c
i
= 2
_
cos
2k
6
, sin
2k
6
_
i = 7, 8, ..., 12; k = i 6
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
x
y
z
Fig. 3.3.8. The lattice velocities of the D2Q13-FHP (multi-speed FHP) lattice.
1 2
3
4 5
6
7 8
9
10 11
12
0
The lattice tensors of rank 2
L
(D2Q13FHP)
= 51
(3.3.28)
and of rank 4
L
(D2Q13FHP)
= 15 (
) (3.3.29)
are isotropic.
3.3.5 Exercises
Exercise 3.3.2. (*)
Prove Theorem 3.3.1 for the special case of tensors of rank 2 in two dimen-
sions. Do reections play any role?
Exercise 3.3.3. (*)
Show that isotropic tensors of the same rank and dimension form a linear
space.
Exercise 3.3.4. (*)
Prove that there are no isotropic tensors of rank 1.
Prove (in 2D) that T
is not isotropic. Note:
is 1 when
all indices are equal and 0 otherwise. Thus it is the generalization of the
Kronecker symbol
.
Prove (in 2D) that T
is isotropic.
3.4 Desperately seeking a lattice for simulations in three dimensions 105
3.4 Desperately seeking a lattice for simulations in three
dimensions
Frisch, Hasslacher and Pomeau (1986) found out that in addition to mass and
momentum conservation lattice-gas cellular automata for the Navier-Stokes
equation must reside on a grid with sucient symmetry. The importance of
this insight can hardly be overestimated. Thus the rst task in the develop-
ment of a LGCA for simulations in 3D is to nd a lattice with appropriate
symmetry. This is not as easy as in 2D.
3.4.1 Three dimensions
In close analogue to 2D where lattice vectors c
i
were dened by the corners of
regular polygons let us dene c
i
in 3D by the corners of regular polytopes. In
three dimensions only a few regular polytopes, namely the Platonic solids
30
,
lead to lattice vectors of equal length. There are exactly ve Platonic solids:
tetrahedron, hexahedron (cube), octahedron, dodecahedron and icosahedron
(compare Table 3.4.1 and Fig. 3.4.1).
The cube and the octahedron are dual to each other in the following sense:
the mid-points of the faces of a cube yield the corners of an octahedron and
the mid-points of an octahedron yield the corners of a cube. In the same
sense dodecahedron and icosahedron are dual to each other. The tetrahedron
is self-dual: its dual solid is also a tetrahedron.
Rotations and reections which transform the solid onto itself form a group
which is referred to as the symmetry group of the solid. Each rotation or
reection which transforms the solid onto itself does the same thing with
the dual solid embedded. Therefore dual regular polytopes show identical
symmetries and their corresponding symmetry groups are isomorphic (Led-
ermann, 1985a). There may be further polytopes with the same symmetry
group which, however, are not regular.
In order to be useful as building blocks of a lattice-gas cellular automata,
a polyhedron must show a large enough symmetry group (this constraint is
fullled only by the dodecahedron and its dual partner the icosahedron) and
in addition must ll the whole space
31
. The cube is the only Platonic solid
whichs completely lls the space without gaps (Ledermann, 1985a). Thus in
3D there is no polyhedron which respects all constraints.
30
Denition of Platonic solids: convex polytopes, bounded by regular congruent
polyhedrons and with equal number of edges meeting at each corner.
31
Actually it is required that all corners (which are nodes of the lattices) are
connected to an equal number of nodes by the lattice vectors c
i
. Gaps between
the polytopes could lead to nodes with a smaller number of nearest neighbors.
As an analog in 2D consider the parqueting of the plane with octagons.
Fig. 3.4.1. The ve Platonic solids.
Tetrahedron
Hexahedron (Cube) Octahedron
Dodecahedron Icosahedron
Table 3.4.1. The ve Platonic solids: number of faces F, edges E and corners C.
Eulers polyhedron theorem: F E + C = 2
polyhedron F E C dual solid
tetrahedron 4 6 4 tetrahedron
hexahedron (cube) 6 12 8 octrahedron
octrahedron 8 12 6 hexahedron (cube)
dodecahedron 12 30 20 icosahedron
icosahedron 20 30 12 dodecahedron
Table 3.4.2. Lattice tensors of ranks 2 to 6 for the ve Platonic solids: isotropic
(+) or not () (from Wolfram, 1986). The number of lattice vectors is equal to the
number of corners C.
polyhedron C 2 3 4 5 6
tetrahedron 4 +
hexahedron (cube) 8 + + +
octrahedron 6 + + +
dodecahedron 20 + + + +
icosahedron 12 + + + +
Historical remark: In Mysterium Cosmographicum (1596) Johannes Kepler
(1571-1630) has suggested that the distances of the then known planets -
Mercury to Saturn - are constrained by the Platonic solids which are alter-
natingly inscribed and circumscribed to spheres. (compare Fig. 3.4.2).
Fig. 3.4.2. In Mysterium Cosmographicum (1596) Johannes Kepler (1571-1630)
has suggested that the distances of the then known planets - Mercury to Saturn -
are constrained by the Platonic solids which are alternatingly inscribed and circum-
scribed to spheres. The six spheres correspond to the six planets, Saturn, Jupiter,
Mars, Earth, Venus, Mercurius, separated in the order by cube, tetrahedron, dodec-
ahedron, octahedron and icosahedron (adapted from Weyl, 1989).
3.4.2 Five and higher dimensions
A further possibility is to nd lattices in higher dimensions with sucient
symmetry. The quantities of interest can then be obtained by appropriate
projections from higher dimensions down to 3D. In ve and higher dimen-
sions there are only three regular polytopes
32
for each dimension, namely
32
These three regular polytopes exist also in lower dimensions.
the simplex
33
, the hypercube
34
and its dual solid. The corresponding lattice
tensors of rank 4 are isotropic only for D < 3 for the simplex and only for
D < 4 for the hypercube and its dual solid.
3.4.3 Four dimensions
The last chance lies in four dimensions where in addition to the simplex, the
hypercube and its dual solid there exist further three regular polytopes which
can be characterized by the Schl ai symbols
35
3, 4, 3, 3, 3, 5, and 5, 3, 3.
The 3, 4, 3-polytop is referred to as face-centered hypercube (FCHC). It has
24 corners with coordinates which are permutations of (1, 1, 0, 0). The
corresponding lattice tensors are isotropic up to 4th rank inclusively. Thus
FCHC is the lattice searched for! The 24 lattice vectors of FCHC read
(1, 1, 0, 0), (1, 0, 1, 0), (1, 0, 0, 1),
(3.4.1)
(0, 1, 1, 0), (0, 1, 0, 1), (0, 0, 1, 1).
The projections of FCHC into 3D space are shown in Figures 3.4.3 and 3.4.4.
There are 12 velocities with c
i4
= 0 (Fig. 3.4.3) and two times 6 velocities
with c
i4
= 1 (Fig. 3.4.4).
The other regular polytopes 3, 3, 5 and 5, 3, 3 are dual to each other. The
3, 3, 5 polytope has 120 corners and the corresponding lattice tensors are
isotropic up to 8th rank inclusively. The 3, 3, 5 and 5, 3, 3 polytopes have
not been mentioned by Wolfram (1986) as alternatives for applications in
lattice-gas automata. One reason at least is simplicity: FCHC has much less
corners than the 3, 3, 5 polytop.
Further reading on regular polytopes: Coxeter (1963).
33
The set of all vectors x which respect the constraints
x =
1
x
1
+ ... +
r
x
r
, x
i
R
D
,
i
R,
i
0,
r
i=1
i
= 1,
are referred to as simplex [x
1
, ..., x
r
]. Despite of some degenerated cases 2, 3,
and 4 points lead to a line segment, a (planar) triangle and a tetrahedron as
simplices (Ledermann, 1985a, p. 108).
34
The hypercube
D
in D dimensions has 2
D
corners e
1
... e
D
. Its 2 D
surfaces are (D1)-dimensional hypercubes.
1
is a line segment,
2
a square,
and
3
a cube (Ledermann, 1985b). e
i
= (0, ..., 0, 1, 0, ..., 0) are the standard base
vectors.
35
Compare Appendix 6.5.
Fig. 3.4.3. The projection (along the 4th axis) of the lattice velocities of the FCHC
lattice for c
i4
= 0.
1
2
3
4
5
6
7
8
9
10
11
12
x
y
z
Fig. 3.4.4. The projection (along the 4th axis) of the lattice velocities of the FCHC
lattice for c
i4
= 1.
1
2
3
4
5
6
x
y
z
Exercise 3.4.1. (***)
Is it possible to build a space-lling lattice from the regular polytop 3, 3, 5
alone?
Exercise 3.4.2. (***)
Prove the polyeder theorem of Leonard Euler (1707-1783): The number C of
corners plus the number F of faces equals the number E of edges plus 2:
C +F = E + 2.
Prove that the Platonic solids are the only regular polyhedrons in 3D. Hint:
Apply the polyeder theorem of Euler.
3.5 FCHC 113
3.5 FCHC
DHumières, Lallemand and Frisch (1986) proposed the face-centered hyper-
cube (FCHC) as a lattice with sucient symmetry for hydrodynamic sim-
ulations in 3D but without specifying collision rules. Those were given for
the rst time by Henon (1987a) and later on modied by Rem and Somers
(1989), Somers and Rem (1989) and others.
The 24 vectors of the FCHC lattice are listed in Section 3.3. The collision
rules have to respect the following constraints:
1. The number of particles is conserved in each collision (conservation of
mass).
2. The momentum is conserved in each collision.
3. There are no conserved quantities except of mass and momentum (no
spurious invariants).
4. The exclusion principle is valid: at each node there sits at most one
particle per lattice velocity (or per cell).
5. The collision rules share the symmetry of the lattice or, in other words,
they are invariant under arbitrary transformations by elements of the
isometric (symmetry) group (.
6. The collisions respect the semi-detailed balance.
The FHP-model with six lattice velocities (FHP-I) has only 2
6
= 64 dierent
states per node. Therefore the collision rules could be derived by hand. On
the contrary, for FCHC there are 2
24
= 16 777 216 dierent states. Thus the
collision rules have to be specied by an automatic algorithm.
3.5.1 Isometric collision rules for FCHC by Henon
Despite the constraints given above there is space for almost unnumerable
many dierent collisions. Therefore Henon (1989a) introduced further con-
straints (referred to as Henon constraints):
1. Every collision is an isometry
36
(isometric collision rules). Motivation:
An isometry is simpler than an arbitrary transformation.
The general constraints 1 and 4 are fullled automatically.
36
Isometries are mappings g which keep the distances d(, ) of arbitrary points
, invariant: d(g(), g()) = d(, ). Rotations and reections are isometries.
All collisions of HPP and FHP (without rest particles) are isometries.
2. The isometry depends only on the total momentum.
3. The isometry which is to be applied will be chosen by random out of the
optimal (with respect to low viscosity, see below) isometries.
Despite the rst two Henon constraints there exists in addition to the identity
at least one nontrivial isometry for each total momentum.
The various isometries contribute dierently to the shear viscosity which
should be kept as low as possible. Henon could quantify these contributions
and thereby classify the optimal isometries. This isometric group ( contains
1152 elements. It can be created by ve generators.
In order to keep the collisions easy to take at a glance, Henon introduced
a kind of normal form of the components of momentum q
1
, q
2
, q
3
, q
4
. Note
that this is not a further constraint to the collision rules. For every state
there exists an isometry which transforms that state into a state (normalized
momenta) with
q
1
q
2
q
3
q
4
and (q
4
= 0 or q
1
+q
4
< q
2
+q
3
). (3.5.1)
Thereby the problem of constructing collision rules for FCHC is simplied
considerably because there exist only 37 dierent normalized momenta.
Each collision proceeds by successive application of three isometries:
1. Transformation of the initial state by an isometry g ( into the state
with the appropriate normalized momentum.
2. Application of an optimal isometry M (proper collision).
3. Transformation with g
1
( back to the original coordinates.
Thus the total transformation reads g
1
M g.
3.5.2 FCHC, computers and modied collision rules
Despite the extreme simplication of the original problem of constructing
collision rules for FCHC by Henon the coding of these rules require the in-
troduction of a very large look-up table that contains the nal state (after
collision) for each initial state (before collision). The necessary memory can
be estimated as follows. The coding of a single state requires 24 bits. On a
CRAY two states can be packed into one word (8 bytes = 64 bits). Thus for
the storage of 2
24
= 16 777 216 initial and the same number of nal states 8
times 16 777 216 bytes or approximately 130 Mbytes are required. This is a
severe obstacle for computers with small memories.
Relatively few applications of the FCHC model have been published. A paper
of Chen et al. (1991c) on the ow through porous media is especially remark-
able. The calculations were performed on a CRAY-YMP (core size of a few
3.5 FCHC 115
hundred Mbytes). The size of the look-up table which actually should contain
16 million entries could be somewhat reduced by exploiting the hole-particle
symmetry. The propagation has been coded in CRAY-Assembler. Despite all
these machine-specic measures the update rates per node are not higher
than those of the PI model (compare Section 3.6) in 3D which is coded in
standard C (Wolf-Gladrow and Vogeler, 1992). For a fair comparison of the
two models, however, the dierent values of the shear viscosities have to be
taken into account.
The collision rules were simplied quite drastically by Rem and Somers (1989)
even risking the violation of the semi-detailed balance. The resulting look-up
table requires only 40 kbytes. Computer experiments show good agreement
with theoretical predictions such as Fermi-Dirac distribution and the shear
viscosity as a function of density. Yet, the shear viscosity is three time higher
than its optimal value.
In addition to his isometric collision rule Henon has proposed a purely ran-
dom rule: The nal state will be randomly chosen out of all states with the
same mass and momentum as the initial state. The values of the shear vis-
cosity for both rules proposed by Henon are comparable (Henon, 1987a).
3.5.3 Isometric rules for HPP and FHP
Henon applied his method for the construction of isometric collision rules
also to the HPP and the FHP model (without restparticles). For HPP one
obtains collision rules which are identical to those of the original formulation:
the head-on collision is the only collision. The model is too simple to allow
additional collisions. On the other hand the Henon constraints are not too
restrictive to forbit all nontrivial transformations.
For the FHP model the isometric collision rules read
(i, i + 3) (i + 1, i + 4) or (i 1, i + 2),
(i, i + 2, i + 4) (i + 1, i + 3, i + 5) or (i, i + 2, i + 4),
(i 1, i, i + 2) (i 2, i, i + 1),
(i + 1, i + 2, i + 4, i + 5) (i, i + 2, i + 3, i + 5) or
(i, i + 1, i + 3, i + 4),
i.e. the two-particle, the three-particle, the two-particle with observer and the
four-particle collisions. The only dierence in relation to the original formu-
lation (compare Section 3.2) occurs for the three-particle collision: according
to the isometric rules the particle velocities will be changed with a probabil-
ity of 1/2 instead of 1 as in the original rules. This can be interpreted as an
indication that the three-particle collision does not contribute much to the
reduction of the shear viscosity. This collision was introduced to destroy a
spurious invariant.
3.5.4 What else?
Implementation of the FCHC model: in the 4th dimension the model en-
compasses only two layers and periodic boundary condition.
The fourth component of the momentum behaves like a passive scalar
(Frisch et al., 1987).
Variants of the FCHC model are listed in Table 3.5.1.
Further reading: Rivet et al. (1988), Shimomura et al. (1988), Henon
(1989), Cancelliere et al. (1990), Dubrulle et al. (1990), Ladd and Frenkel
(1990), Vergassola et al. (1990), Rivet (1991), Benzi et al. (1992), Henon
(1992), van der Hoef et al. (1992), Verheggen (1992), van Coevorden et al.
(1994), Adler et al. (1995).
Classify the 64 dierent states at each node of the FHP-I model according
to particle number and momentum.
How many dierent 3-momenta can be realized in the FCHC model?
Concluding remark: If one restricts oneself to models with a single lattice
speed there exists only the FCHC model for hydrodynamic simulations in
3D. The collision rules of the FCHC model are much more complicated than
those of the FHP model. Later on we will discuss multi-speed models as an
alternative to FCHC.
3.5 FCHC 117
Table 3.5.1. Variants of the FCHC model (compare Dubrulle et al., 1990). Roughly,
the Reynolds coecient R
max
measures the inverse viscosity in lattice units.

Name Rest Semi-detailed R
max
R
max
References
particles balance Boltzmann measured
FCHC-1 0 Yes 2.00 2.0 Henon (1987), Rivet (1987)
FCHC-2 0 Yes 6.44 - Henon (1989)
FCHC-3 0 Yes 7.13 6.4 Henon (1989), Rivet (1988a,b)
FCHC-4 0 Yes 7.57 - Henon (1989)
FCHC-5 3 Yes 10.71 - Henon (1989)
FCHC-6 0 No 17.2 - Dubrulle (1988)
FCHC-7 3 No () 7.9 Dubrulle et al. (1990)
FCHC-8 3 No 99.7 13.5 Dubrulle et al. (1990)
3.6 The pair interaction (PI) lattice-gas cellular
automata
The FCHC model for hydrodynamic simulations in 3D has complicated colli-
sion rules and thus requires large look-up tables. Extension to problems with
a free surface or to magneto-hydrodynamics seems to be extremely involved.
In 1989 Nasilowski proposed a lattice-gas cellular automata which runs in
2D over a square lattice as well as in 3D over a cubic grid. In contrast to
FHP or FCHC the state of a cell of the PI model is characterized by D + 1
(D spatial dimension) bits instead of only one. The interaction
37
at a node
consists of a succession of interactions between pairs of cells (thus the name
PI = pair interaction). The splitting into pair interactions allows an ecient
coding with bit-operators (C) or bit-functions (FORTRAN) also in 3D.
A complete discussion of the pair interaction lattice-gas cellular automata
has been given by Nasilowski (1991). Here only the main ideas of the PI-
approach will be explained and consequentely the presentation is restricted
to the two-dimensional case. The extension to 3D is straightforward.
3.6.1 Lattice, cells, and interaction in 2D
The PI model in 2D is based on the square lattice (compare Fig. 3.6.1). As
usual the development in time proceeds by an alternating sequence of local
interaction (only cells of a single node are involved) and propagation to the
nearest neighbor nodes. The lattice splits into two sub-lattices
38
:
At even time levels the particles reside on nodes with even indices (white
circles).
At odd time levels the particles reside on nodes with odd indices (black
circles).
As for HPP, FHP and FCHC there is a cell on each link at each node. The
state of a cell of the PI model is characterized by D+1 bits n
J
(J = 0, 1, ..., D)
where D is the dimension: n
0
is called the mass bit and n
j
(j = 1, ..., D) are
the momentum bits. By convention in this section uppercase indices run from
0 to D whereas lowercase indices run from 1 to D. The momentum bits are
subject to the constraint
37
We speak of interaction instead of collision because some of the rules of the PI
lattice-gas cellular automata cannot be described as collisions between particles.
38
Points with a combination of an even and an odd index will never be occupied
by particles. Therefore they will not be called nodes and they are not shown in
Fig. 3.6.1.
3.6 The pair interaction (PI) lattice-gas cellular automata 119
n
j
n
0
(3.6.1)
which can be interpreted as the momentum of empty cells vanishes. The
vector linking two neighboring nodes is termed c (lattice velocity). The com-
ponents of c take on the values 1 or 1. The links to the neighboring nodes
and the corresponding cells of a node will be labelled a, b, c, d (compare
Fig. 3.6.2). The corresponding lattice velocities are given by c
a
= (1, 1),
c
b
= (1, 1), c
c
= (1, 1), c
d
= (1, 1). The momentum m is dened
component-wise:
m
j
:= n
j
v
j
, j
= 1, ..., D (3.6.2)
(remark: no summation convention here!). This denition is rather unusual
because in general the momentum does not point to the same direction as
the velocity. This can be illustrated by considering all possible states of cell
a:
1. Mass bit n
0
= 0 all momentum bits vanish (according to eq. 3.6.2):
the cell is empty.
2. Mass bit n
0
= 1, all momentum bits vanish: rest particle.
3. n
0
= n
1
= 1, n
2
= 0: particle with x-momentum only.
4. n
0
= n
2
= 1, n
1
= 0: particle with y-momentum only.
5. n
0
= n
1
= n
2
= 1: particle with momentum in the diagonal direction;
this is the only case where m and c point to the same direction.
Fig. 3.6.1. The sub-lattices of the PI lattice-gas cellular automata in 2D

g g g g
g g g g
Z
Z
`
``
Z
Z
`
`
What was Nasilowskis motivation to introduce the somewhat strange de-
nition of the momentum? There are at least two good reasons:
1. The component-wise denition of the momentum allows a splitting of the
interaction into pair interactions (see below).
Fig. 3.6.2. Structure of the nodes and cells of the PI lattice-gas cellular automata.
g
Z
Z
Z
'
_
_
_
`
`
``
'
_
_
_
`
`
`
'
_
_
_
Z
Z
Z
'
_
_
_
a
b
c
d
Z
Z
Z
Z
Z
Z
Z
`
_
n
0
`
_
n
1
`
_
n
2
c
a
2. The symmetry of the square lattice is not sucient to assure the isotropy
of 4th rank lattice tensors (compare HPP and Section 3.3 on lattice
tensors). The component-wise denition of the momentum introduces a
new degree of freedom in that the momentum can uctuate with respect
to the direction of velocity. This freedom may open a new route to
isotropy.
The interaction is composed of the following sequence of pair interactions
(compare Fig. 3.6.3):
1. Interaction in x-direction between the cells a and b and between the cells
c and d
2. followed by interaction in y-direction between the cells a and d and be-
tween the cells b and c.
Fig. 3.6.3. PI: interaction between pairs of cell rst in x-, then in y-direction.
x-interaction
Z
Z
Z
Z
`
`
`
``
Z
Z
Z
Z
`
`
`
`
y-interaction
Z
Z
Z
Z
`
`
`
``
Z
Z
Z
Z
`
`
`
`
The pair interaction rules are designed according to the maxim whatever is
not forbidden is allowed. Nasilowski (1991) formulates three constraints:
1. The interaction must conserve mass and momentum (as all good lattice-
gas cellular automata should do).
2. The interaction must be reversible, i.e. the mapping between initial and
nal state is one-to-one. This allows the calculation of the statistical
equilibrium by applying the Gibbs formalism.
3. The interaction should yield a maximal change of the state of a node. The
identity fullls the rst two constraints but leads to spurious (additional)
invariants as, for example, the particle number on each diagonal of the
lattice. Such invariants will produce deviations from the hydrodynamic
behavior in the macroscopic limit and therefore should be avoided.
The rules given by Nasilowski (1991, p. 107) obey the rst and second con-
straint and encompass all allowed pair interactions (compare also Fig. 3.6.4)
which most probably lead to the maximal change.
3.6.2 Macroscopic equations
The rather lengthy calculations of the equilibrium distribution and the multi-
scale expansion has been given by Nasilowski (1991). The rst order terms
of a multi-scale expansion and the rescaling of certain quantities leads for
0
= 1/2 to the continuity equation
u = 0 (3.6.3)
and the Euler equation
(
t
+u)u +P = 0 (3.6.4)
where
P =
p
0
=
4
9
(3.6.5)
is the kinematic pressure. The hydrodynamic velocity u is related to the
momentum density q by
u =
8
9
q (3.6.6)
where the hyper-momentum density q is dened component-wise by
q
J
= 2
D
v
v
J
n
v J
(3.6.7)
with the hyper-velocity v = (v
0
, v
1
, ..., v
D
) = (1, c). In particular, q :=
(q
1
, ..., q
D
) is the momentum density, and := q
0
is the mass density. The
viscosity resulting from the second order terms of the multiscale expansion
is anisotropic: a tensor of 4th order instead of a scalar.
Fig. 3.6.4. Pair interaction in horizontal (x-) direction: all possible congurations
and changes. Open cycles denote empty cells, lled cycles denote occupied cells, and
arrows indicate the momentum.
1.)
'
`
``
_
_

'
`
``
2.) _
_
_
_
3.)
Z
Z
'
_
_
_

Z
Z
'
_
4.)
'
_
'
_
_
_
'
_
'
_
5.) _
'
_
_
_
_
'
_
6.)
'
'
_
_

Z
Z
`
``
7.)
Z
Z
_
_
_

'
8.)
'
_
_
_

`
``
9.)
_
_
_ _
10.)
'
'
_
_
_
'
_
'
11.)
'
_
_
_
'
_

Fig. 3.6.5. Simulation with PI-LGA of a Karman vortex street in 2D at a Reynolds
number of 80 (Wolf-Gladrow et al., 1991): ow past a plate with upstream on the
left. The gure shows the perturbation of the velocity after 80, 000 time steps. The
homogeneous ow eld was subtracted to make the eddies clearly visible. The lattice
consists of 6400 times 3200 nodes.
3.6.3 Comparison of PI with FHP and FCHC
The PI-LGCA has some disadvantages compared with FHP and FCHC. PI
can only be applied at the particular mass density
0
= 1/2 and the vis-
cosity is non-isotropic. On the other hand, the advantages are impressive.
The pressure does not depend explicitly on the velocity and simulations in
3D are possible with portable code (Wolf-Gladrow and Vogeler, 1992). As an
example Vogeler and Wolf-Gladrow (1993) have calculated drag coecients
for ows past obstacles in two and three dimensions.
Exercise 3.6.1. (*)
PI in 2D and 3D: How many dierent states are possible at a single node?
How large is the number of states on a lattice with only four nodes?
Does the result of the interaction depend on the order of the pair interactions?
Find nontrivial rules for a pair interaction between the cell a and c and
between b and d (diagonal pair interaction).
3.6.4 The collision operator and propagation in C and FORTRAN
C:
IXM = 32; IXM1 = IXM - 1;
IYM = 1024; IYM1 = IYM - 1;
LAST = LENGTH - 1; /* the last bit */
/* ----- interaction on 1. sub-lattice ----- */
/* ----- interaction: x-direction ----- */
/* ----- pair a <--> b ----- */
for(ix=0; ix < IXM; ix++)
for(iy=1; iy < IYM1; iy++) {
ab0 = a10[ix][iy] & b10[ix][iy];
ba0 = a10[ix][iy] b10[ix][iy];
nab1 = (a11[ix][iy] | b11[ix][iy]);
chab0 = ba0 & nab1;
chab1 = ab0 & (a11[ix][iy] b11[ix][iy]);
chab2 = ((ba0 & nab1) | ab0) & (a12[ix][iy] b12[ix][iy]);
/* array_1 --> array_2 */
a20[ix][iy] = a10[ix][iy] chab0;
b20[ix][iy] = b10[ix][iy] chab0;
b21[ix][iy] = b11[ix][iy] chab1;
b22[ix][iy] = b12[ix][iy] chab2; }
/* ... interactions c <-> d, a <-> d, b <-> c ... */
/* propagation: */
/* --- a-direction --- */
for(ix=0; ix < IXM; ix++)
for(iy=0; iy < IYM; iy++) {
a10[ix][iy] = a20[ix][iy];
a11[ix][iy] = a21[ix][iy];
a12[ix][iy] = a22[ix][iy]; }
/* --- b-direction --- */
for(ix=0; ix < IXM1; ix++)
for(iy=0; iy < IYM; iy++) {
b10[ix][iy] = (b20[ix][iy]>>1) + (b20[ix+1][iy]<<LAST);
b11[ix][iy] = (b21[ix][iy]>>1) + (b21[ix+1][iy]<<LAST);
b12[ix][iy] = (b22[ix][iy]>>1) + (b22[ix+1][iy]<<LAST); }
FORTRAN:
IXM = 32
IXM1 = IXM - 1
IYM = 1024
IYM1 = IYM - 1
c last bit
LAST = LENGTH - 1
c ----- interaction on 1. sub-lattice -----
c ----- interaction: x-direction -----
c ------------------ pair (a,b)
DO iy=1,IYM
DO ix=1,IXM
ab0 = iand(a10(ix,iy),b10(ix,iy))
ba0 = ieor(a10(ix,iy),b10(ix,iy))
nab1 = not(ior(a11(ix,iy),b11(ix,iy)))
c --- change ?
chab0 = iand(ba0,iand(nab1,sb1(ix,iy)))
chab1 = iand(ab0,not(iand(ieor(a11(ix,iy),
1 b11(ix,iy)),sb1(ix,iy))))
chab2 = iand(ior(iand(ba0,nab1),ab0),
1 iand(ieor(a12(ix,iy),b12(ix,iy)),sb1(ix,iy)))
c --- set new values
a20(ix,iy) = ieor(a10(ix,iy),chab0)
b20(ix,iy) = ieor(b10(ix,iy),chab0)
ENDDO
ENDDO
c ... interactions c <-> d, a <-> d, b <-> c ...
c propagation:
c --- a-direction ---
DO iy=1,IYM
DO ix=1,IXM
a10(ix,iy) = a20(ix,iy)
ENDDO
ENDDO
c --- b-direction ---
DO iy=1,IYM
DO ix=1,ixm1
b10(ix,iy) = ishft(b20(ix,iy),-1)
1 + ishft(b20(ix+1,iy),LAST)
ENDDO
ENDDO
3.7 Multi-speed and thermal lattice-gas cellular
automata
The macroscopic equations derived from the microdynamics of lattice-gas
cellular automata include 4th rank tensors which are composed of the lattice
velocities and therefore are referred to as lattice tensors (compare Section
3.3 and especially Eq. (3.3.4)). The isotropy of these tensors, which depend
on the symmetry of the underlying lattice, is an essential condition to ob-
tain the Navier-Stokes equations. Hasslacher (1987) has shown that models
with several dierent non-vanishing lattice speeds may be equivalent to mod-
els with a single speed but larger symmetry group. The resulting 4th rank
tensors include speed dependent weights and are referred to as generalized
lattice tensors. By appropriate choice of the weights these tensors can be
isotropic for relative low symmetry of the lattice. In contrast to single speed
models the pressure in multi-speed models does not depend explicitly on the
ow velocity. The collision rules can be chosen such that in addition to mass
and momentum also kinetic energy will be conserved. Such models are called
thermal LGCA.
Models with several dierent speeds have been encountered before, namely
the FHP variants with rest particles (FHP-II, FHP-III). Particles with van-
ishing speed have no inuence on the isotropy of the lattice tensors. Therefore
only models with dierent non-vanishing speeds are called multi-speed mod-
els.
3.7.1 The D3Q19 model
DHumières, Lallemand, and Frisch proposed already in 1986 - together with
FCHC - a multi-speed model in 3D with 19 lattice velocities (Eq. 3.3.24) and
three speeds (0, 1, and

2). The collisions shall conserve mass , momentum
j and kinetic energy density
K
which are dened as follows
=
I
N
I
j = u =
I
N
I
c
I
K
=
I
N
I
c
2
I
2
where the index I = (, i) indicate the speed ( = 0, 1, 2 for the rest particles
and particles with speed 1 and

2, respectively) as well as the direction i.
3.7 Multi-speed and thermal lattice-gas cellular automata 129
The exclusion principle and semi-detailed balance lead to Fermi-Dirac distri-
butions for the mean occupation numbers
N
I
= f
FD
(Q
I
) =
1
1 + exp(Q
I
)
whereby
Q
I
= +c
I
u +c
2
I
is a linear superposition of the collision invariants. The Lagrange multipliers
, , and can be calculated by expansion for small Mach numbers (compare
analogous calculations for FHP).
The D3Q19 model is the rst lattice-gas cellular automata with an energy
term in the distribution function. This raises the question why similar terms
do not occur in the other models (FHP, FCHC, PI).
For FHP without rest particles mass and kinetic energy density are iden-
tical:
2
K
=
i
N
i
c
2
i
..
=1
=
i
N
i
=
FHP with rest particles: the collisions including rest particles do not con-
serve kinetic energy.
PI: the kinetic energy density can be dened by
K
:=
i
p
2
i
2m
(particle mass m = 1); some of the interactions do not conserve energy.
FCHC: same as for FHP, i.e. the kinetic energy density is proportional to
the mass density for the version without rest particles and the collisions
including rest particles do not conserve kinetic energy.
The advection term of the macroscopic momentum equation of the D3Q19
model contains the following 4th rank tensor
T
=

2
0
2
I
d
(1 d
)(1 2d
)
. .
= w
c
I
c
I
c
I
c
I
where d
are the equilibrium occupation numbers at vanishing ow speed

(the zeroth order term of the Taylor expansion) which depend only on the
lattice speed and not on direction. The new feature - compared to single
speed models - is the occurrence of the weights
w
= d
(1 d
)(1 2d
)
which can inuence the transformation properties (isotropy) of T
(of
course the coecient
2
0
/2 does not play a role). The tensor T
of the
D3Q19 model is isotropic if
d
1
(1 d
1
)(1 2d
1
) = 4d
2
(1 d
2
)(1 2d
2
) (3.7.1)
(dHumières et al., 1986; compare also Section 3.3), i.e. the occupation num-
bers of speed-1 and speed-
2 particles must respect a certain ratio. In other

words, the tensor is isotropic only at a certain temperature where the right
number of high energy states are excited. For small densities (d
1) rela-
tion (3.7.1) can be approximated by d
1
= 4d
2
, i.e. the occupation numbers
of cells with speed 1 must be four times as high as for cells with speed

2
(compare Fig. 3.7.1 for the ratio of d
2
/d
1
for nite values of d
1
). Thus a
certain non-isotropy of the occupation numbers ensures the isotropy of the
tensor T
.
Fig. 3.7.1. The gure shows one solution of the cubic equation d
1
(1d
1
)(12d
1
) =
4d
2
(1 d
2
)(1 2d
2
) (solid line). For small values of d
1
it can be approximated by
d
1
= 4d
2
(broken line).
0 0.1 0.2 0.3 0.4 0.5
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
d
1
4

d
2
Exercise 3.7.1. (*)
Which interactions of the PI model violate energy conservation?
Exercise 3.7.2. (***)
Propose collision rules for D3Q19.
Chen et al. (1989) reduced the multi-speed model of dHumières et al. (1986)
to 2D. They proposed collision rules and performed some numerical experi-
ments. The model encompasses 9 lattice velocities:
c
0
c
1,2
= (1, 0) 1-particles
c
3,4
= (0, 1) 1-particles
c
5,6,7,8
= (1, 1)

2-particles.
All particles have the same mass. Only collisions between two particles are
considered (compare Fig. 3.7.2):
1. Head-on collision of two particles with speed c = 1: as for HPP both
particles are rotated (in the same sense) by 90
.
2. Head-on collision of two particles with speed c =

2: as for HPP both
particles are rotated (in the same sense) by 90
.
3. Collision between a

2-particle and a rest particle: two particles with
speed c = 1 leave the node under 45
with respect to the incoming
2-particle. The inverse process is also allowed.

4. Collision between a

2-particle and a particle with speed c = 1 from
dierent quadrants: the

2-particle will be rotated by 90
whereas the
velocity of the 1-particle will be reversed (identical to rotation by 180
).
All these two-particle collisions conserve mass, momentum, and kinetic en-
ergy. The third type of collisions changes the number of particles with given
speed. The tensor of 4th rank in the advection term reads
T
= const. (
)
+2 [d
1
(1 d
1
)(1 2d
1
) 4d
2
(1 d
2
)(1 2d
2
)]
.
The second part of the term, which destroys isotropy, vanishes under the same
condition (3.7.1) given by dHumières et al. (1986). In the limit of small occu-
pation numbers one obtains d
1
/d
4
= 4 and a kinetic energy density
K
= 1/3.
In the same limit the pressure does not depend explicitly on the ow velocity.
For nite mass densities, however, there occurs an u
2
term as in FHP.
Furthermore, the model allows simulation of pure heat conduction problems
(u = 0). However, this is possible with simpler lattice-gas cellular automata
(see, for example, Chopard and Droz, 1988 and 1991) or with classical meth-
ods like nite dierences. Biggs and Humby (1998) summarize their discussion
of thermal LGA as follows:
Thermal LGA has not been applied to problems of any real signicance.
The only simulation of note is that of Chen et al. (1991a) who briey consid-
ered the Benard convection problem. The relative paucity of non-isothermal
LGA studies is perhaps not surprising given the relative immaturity of this
level of LGA model. The models have, however, developed to a point where
serious application can be contemplated provided substantial validation work
is undertaken.
Show that the collision rules of Chen et al. (1989) contain all possible two-
particle collisions.
Find three-particle collisions which conserve mass, momentum, and energy.
1.
2.
3.
4.
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
'
'
'
,_
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
`
`
`
`
`
``
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
'
'
_ _
_
_
_
_
_
_
_
_
_
Fig. 3.7.2. Collision rules proposed by Chen et al. (1989) for the D2Q9 model. The
rest-particle in rule three is indicated by a circle.
Fahner (1991) proposed a model with 21 lattice velocities in 2D (compare
Section 3.3 and especially Eq. (3.3.18)). In addition to the 9 velocities of the
D2Q9 model Fahner included the speeds 2 and

5. The number of possible
states per node of 2
21
= 2 097 152 is almost comparable with that of FCHC.
Fahner proposed the following collision rule which reminds on Henons ran-
dom rule for FCHC:
At each node mass, momentum, and kinetic energy will be calculated (these
quantities characterize the macrostate of each node).
The nal microstate of each node will be chosen among all microstates
which are compatible with its macrostate.
To keep the number of collision states low and thus the look-up table small,
Fahner took into account only collisions with a maximum of ve particles
involved. The number of these collisions is maximized by initializing a mean
mass density with two particles per node.
The equilibrium occupation numbers obey a Fermi-Dirac distribution
N
i
=
1
1 + exp( +
i
qc
i
)
(3.7.2)
where c
i
are the lattice velocities,
i
= c
2
i
0, 1, 2, 3, 4, 5 are single-particle
energies and q, , are Lagrange multipliers. The Lagrange multipliers can
be calculated by expansion for small Mach numbers. Explicit expressions can
be found in Fahner (1991).
3.7.4 Transsonic and supersonic ows: D2Q25, D2Q57, D2Q129
Kornreich and Scalo (1993) proposed multi-speed models similar to that of
Fahner (1991) but with 25, 57, and 129 lattice velocities in 2D. The Lagrange
multipliers were calculated numerically instead of applying the usual expan-
sion for small Mach numbers. The resulting numerical distribution functions
were used to determine the pressure tensor. Kornreich and Scalo found that
the velocity dependent term in the expression for the pressure does not in-
crease as fast as it can be predicted by applying the expanded distributions.
On the contrary, this term decreases again when the (macroscopic) ow veloc-
ity u is near one of the (microscopic) lattice velocities c
i
. Thus, in principle
simulations of transsonic and supersonic ows are possible. These models
are not without problems. The collision rules are complicated already for
the D2Q25 model despite the fact that Kornreich and Scalo considered only
two-particle collisions. The memory demand is enormous and the viscosity is
larger than for FHP.
3.8 Zanetti (staggered) invariants 135
3.8 Zanetti (staggered) invariants
In addition to local - at one node, at one time level - invariants there may
exist some invariants which involve quantities at dierent nodes and dierent
time levels. Zanetti (1989, p. 1539) gives a simple example:
The presence of these new invariants can be easily understood by using a
trivial one-dimensional example. Let g(x) be the linear momentum of the par-
ticles present at site x, dene G
e
(t) =

x,even
g(x, t), G
o
(t) =

x,odd
g(x, t)
as the total momentum of the particles on even or odd sites, and let the col-
lision rules conserve the momentum and the number of particles at each site.
Since the particles can only hop between nearest neighbors, G
e
and G
o
are
exchanged at each time step. The dynamics of this one-dimensional model
allows three conserved quantities: M, G
e
+G
o
, and H = (1)
t
(G
e
G
o
). The
rst two are the usual number of particles and the total linear momentum;
the third is due to our extremely simplied dynamics.
3.8.1 FHP
Zanetti (1989) has found three
39
staggered invariants of the FHP-III model:
H
i
= (1)
t
r
(1)
b
i
r
c
i
j; i = 1, 2, 3, (3.8.1)
where c
i
is obtained by rotating c
i
by /2 counterclockwise, b
i
= (2/
3)c
i
is the reciprocal space vector perpendicular to c
i
and j =

k
c
k
n
k
(x, t) is
the microscopic momentum density.
According to Zanetti the validity of (3.8.1) can be veried by inspection.
More precisely it may be shown (see Exercise 3.8.1) that the H
i
are conserved
by the combination of any single collision plus propagation and for collision-
less propagation. Because the H
i
are linear in the n
k
this provides a complete
proof of the invariance. The Zanetti invariants are also valid for FHP-I and
FHP-II which follows as a direct consequence of the proof sketched above.
3.8.2 Signicance of the Zanetti invariants
Additional invariants further restrict the dynamics by changing the equi-
librium distributions which are functions of all invariants. Fortunately, the
procedures usually applied for initialization create very small values of H
k
(noise). Consequently, in the simulation reported so far there is no clear
indication of the presence of the staggered-momentum density h
i
(r, t) =
(1)
t
(1)
b
i
r
c
i
j (i = 1, 2, 3). But pathological initial conditions as, for
39
Note: H
i+3
= H
i
for i = 1, 2, 3
example, = const, j = 0, h
1
= 0 = h
3
and h
2
= h
0
sin(2y/W)
lead to an exitation of an oscillatory mode in j
x
due to a coupling of the
hydrodynamic modes with the additional (kinetic) modes of the lattice gas
(Zanetti, 1989).
Further reading: Kadano et al. (1989), dHumières et al. (1989, 1990),
Bernardin (1992), Qian (1997). Ernst (1991) calculates the contribution of
the unphysical modes to the stress tensor for FHP.
Consider a lattice with 3 times 3 sites which is empty except for the site
(2, 2) in the middle of the domain. Calculate the Zanetti invariants 1.) with
the initial (t = 0) congurations of Fig. 3.2.2 at site (2, 2) and 2.) after
collision and propagation (t = 1). In addition prove that H
i
are conserved by
a single propagating particle.
Qian et al. (1992) proposed a LGCA for diusion in 1D with four velocities
c
i
(D1Q4) and two dierent masses m
i
:
c
i
= 2c, 2c, c, c
m
i
= m, m, 2m, 2m
There is only one possible collision which conserves mass, momentum and
energy: the head-on collision between a fast (hot) particle with mass m and a
slow (cold) particle with mass 2m which leads to a reversal of their velocities.
For a domain with L sites where L is divisible by four and periodic boundary
conditions nd three spurious invariants staggered in space and time and one
additional spurious invariant staggered in space only.
Can you nd staggered invariants for HPP?
Exercise 3.8.4. (***)
Try to nd staggered invariants for PI.
Exercise 3.8.5. (***)
Try to nd staggered invariants for D2Q9.
Exercise 3.8.6. (***)
Modify the usual procedure to initialize a given distribution of , j and
staggered-momentum densities h
i
(the microscopic densities corresponding
to H
i
).
3.9 Lattice-gas cellular automata: What else? 137
3.9 Lattice-gas cellular automata: What else?
Further reading:
Textbooks: Rothman and Zaleski (1997), Rivet and Boon (to be published).
Proceedings and reviews:
Proceedings of the workshop on Large Nonlinear Systems, Complex Sys-
tems, Vol. 1, No. 4, 1987.
Monaco (1989)
Doolen (1990)
Ernst (1991)
Chen et al. (1995)
Boon et al. (1996)
Biggs and Humby (1998)
Bibliographies: Doolen (1990) and J. Stat. Phys. 68 (3/4), 611-667, 1992.
Further topics:
Mode-mode coupling, long time-tails, divergence of transport coecients
in 2D: Kadano et al. (1989), Frenkel and Ernst (1989), dHumières et al.
(1989), McNamara (1990), van der Hoef and Frenkel (1990), Ernst (1991).
Flows in 3D: Bernsdorf et al. (1999), van Genabeek and Rothman (1999).
Flow through porous media: Balasubramanian et al. (1987), Rothman
(1988), Kohring (1991a,b,c), Knackstedt et al. (1993), Gutfraind et al.
(1995), van Genabeek and Rothman (1996), Koponen et al. (1996, 1997),
Krafczyk et al. (1998), Matsukama (1998), Matsukuma et al. (1998), Ni-
imura (1998), Waite et al. (1998).
Multiphase ows (colored models) in 2D: Rothman and Keller (1988),
Stockman et al. (1990), Gunstensen and Rothman (1991), Kougias (1993),
Rothman and Zaleski (1994, review article), Emerton et al. (1997), Mat-
sukama (1998), Peng and Ohta (1997), Stockman et al. (1997), Tsumaya
and Ohashi (1997), Weig et al. (1997), Ebihara et al. (1998), Sehgal et al.
(1999).
Multiphase ows in 3D: Rem and Somers (1989), Olson and Rothman
(1995, 1997).
Flow of granular media: Karolyi et al. (1998), Manna and Khakhar (1998).
Flow in dynamical geometry: Hasslacher and Meyer (1998)
Poisson solver: Chen, Matthaeus and Klein (1990).
Magnetohydrodynamics (MHD): Chen and Matthaeus (1987), Hatori and
Montgomery (1987), Montgomery and Doolen (1987), Chen et al. (1991),
Succi et al. (1991), Chen et al. (1992), Martinez et al. (1994), Isliker et al.
(1998), Takalo et al. (1999).
Relativistic ows: Balazs et al. (1999).
Chemical reactions, reaction-diusion equations: Weimar et al. (1992),
Weimar (1997), Decker and Jeulin (1997), Vanag and Nicolis (1999).
Burgers equation: Boghosian and Levermore (1987), Cheng et al. (1991),
Nishinari and Takahashi (1998).
Generalized semi-detailed balance condition: Chen (1995, 1997).
Beyond the Boltzmann approximation: Boghosian (1995).
Pattern formation: Chen et al. (1995), Bussemaker (1996), Deutsch and
Lawniczak (1999).
Integer lattice gases, Digital Physics: Boghosian et al. (1997), Chen, Teix-
eira and Molvig (1997), Teixeira (1997).
Quantum mechanics: Boghosian and Taylor (1997), Boghosian and Taylor
(1998), Succi (1998), Yepez (1998).
Further reading: Kohring (1992a,b), Hashimoto and Ohashi (1997), Suarez
and Boon (1997), Tribel and Boon (1997), Buick (1998), Hashimoto et
al. (1998), Lahaie and Grasso (1998), Masselot and Chopard (1998b),
Nicodemi (1998), Tsujimoto and Hirota (1998), Ujita et al. (1998).

Chap 03

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Chap 03

Uploaded by

Copyright:

Available Formats

3.

Lattice-gas cellular automata

3.1 The HPP lattice-gas cellular automata 43

. Mass and momentum are conserved.

(hexagonal symmetry) about an axis

to the left or right) while

. The rst three moments of the lattice velocities c

) form a 64 64 transition matrix A

(r)) P (t, s(.)) . (3.2.23)

). For it to vanish, one must have

where the last two scales will be considered as continuous

) = 0 (continuity equation) (3.2.52)

2. It can be shown that invariance under

as a function of the density per cell d for

0.387 1.08 2.22

for HPP from

. In the next propagation step they will

are dened such that their mean values

2 (in 300 time

3.2 The FHP lattice-gas cellular automata 89

90 3. Lattice-gas cellular automata

and all tensors that result from cyclic per-

is 1 if all indices are

are the cartesian components of the lattice velocities c

(compare Eq. 3.2.55)

is isotropic. The lattice tensor of rank 4 with the following non-vanishing

is isotropic whereas the lattice tensor of rank 4

is non-isotropic. The weights w

is isotropic whereas the lattice tensor of rank 4

is non-isotropic. The weights w

is isotropic whereas the lattice tensor of rank 4

is non-isotropic. The weights w

is isotropic whereas the lattice tensor of rank 4

is non-isotropic. The weights w

is not isotropic. Note:

measures the inverse viscosity in lattice units.

3.6 The pair interaction (PI) lattice-gas cellular automata 123

are the equilibrium occupation numbers at vanishing ow speed

2 particles must respect a certain ratio. In other

with respect to the incoming

2-particle. The inverse process is also allowed.

You might also like